Indonesian J our nal of Electrical Engineering and Computer Science V ol. 42, No. 1, April 2026, pp. 131 ∼ 141 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v42.i1.pp131-141 ❒ 131 T r ophallactic optimization algorithm with mark o v random eld r enement f or str ok e lesion segmentation Hay et Berk ok, Karima Kies, Nac ´ era Benamrane Department of Computer Science, F aculty of Mathematics and Computer Science, Uni v ersity of Science and T echnology Mohamed Boudiaf, Oran, Algeria Article Inf o Article history: Recei v ed Dec 18, 2025 Re vised Feb 7, 2026 Accepted Mar 4, 2026 K eyw ords: Articial bee colon y Computed tomograph y Mark o v random elds Strok e lesion se gmentation T rophallactic optimization algorithm ABSTRA CT Cerebro v ascular accidents (strok es) represent a critical medical emer genc y re- quiring rapid and accurate diagnosis. Automated se gmentation of strok e lesions from computed tomograph y (CT) images remains challenging due t o lo w con- trast, image noise, and high anatomical v ariabilit y between ischemic and hem- orrhagic subtypes. This paper introduces a no v el h ybrid approach combining the trophallactic optimization algorithm (T O A), inspired by cooperati v e nectar e xchange in bee colonies, with mark o v random elds (MRF) for spatial coher - ence modeling. The proposed T O A-MRF method operates semi-automatically from a single user -dened seed point, le v eraging bio-inspired collecti v e intel- ligence to progressi v ely e xplore and rene re gions of interest. The algorithm simulates the enzymatic transformation of nectar into hone y through iterati v e information e xchange between virtual bees, follo wed by MRF-based re gulariza- tion to ensure anatomical consistenc y . Ev aluated on a clinical CT dataset from [Hospital Name], the method achie v es a Dice similarity coef cient of 87.3% for ischemic strok es and 91.2% for hemorrhagic strok es, with an o v erall detection accurac y e xceeding 89%. Comparati v e analysis demonstrates the complemen- tary strengths of T O A e xploration and MRF renement, of fering a rob ust and ef cient solution for clinical strok e assessment with minimal user interv ention. This is an open access article under the CC BY -SA license . Corresponding A uthor: Karima Kies Department of Computer Science, F aculty of Mathematics and Computer Science Uni v ersity of Science and T echnology Mohamed Boudiaf Oran, Algeria Email: karima.kies@uni v-usto.dz 1. INTR ODUCTION Strok e represents one of the leading causes of mortality and morbidity w orldwide, with approxima tely 15 million people af fected annually according to the W orld Health Or g anization (WHO). Early and accurate detection, follo wed by proper classication between ischemi c and hemorrhagic forms, is essential to ensure rapid therapeutic interv ention and impro v e patient outcomes. The critical “golden hour” follo wing strok e onset necessitates immediate medical imaging analysis, typically through computed tomograph y (CT) scans, to guide treatment decisions such as thrombolytic therap y for ischemic strok es or sur gical interv ention for hemorrhagic cases. Ho we v er , automated se gmentation of strok e lesions from CT im ages remains a signicant te chnical challenge. The comple xity arises from multiple f actors: lo w tissue contrast bet ween pathological and health y re gions, presence of imaging artif acts and noise, high inter -patient anatomical v ariability , and subtle intensity J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
132 ❒ ISSN: 2502-4752 dif ferences between dif ferent strok e subtypes. Classical se gmentation approaches based on intensity threshold- ing, acti v e contours, or re gion gro wing methods often demonstrate limited rob ustness, suf fering from sensiti vity to initialization parameters and dif culty handling the heterogeneous nature of cerebro v ascular lesions. T o o v ercome these limitations, bio-inspired optimization approaches ha v e g ained considerable atten- tion in medical image analysis. These methods , inspired by collecti v e beha viors observ ed in natural systems such as ant colonies, bird ocking, or bee sw arms, e xploit distrib uted intelligence mechanisms to ef ciently e xplore comple x solution space s. Among these, the articial bee colon y (ABC) algorithm has demonstrated particular promise due to its balanced e xploration-e xploitation strate gy through cooperati v e interactions be- tween dif ferent bee roles: scouts for e xploration, foragers for local e xploitation, and onlook ers for information sharing. Complementary to bio-inspired optimization, mark o v random elds (MRF) pro vide a po werful proba- bilistic frame w ork for modeling spatial dependencies in image se gmentation tasks. MRFs enable re gularization of se gmentation results by imposing local coherence constraints between neighboring re gions while preserving signicant discontinuities at lesion boundaries, making them particularly suitable for medical imaging applica- tions where anatomical structure and spatial continuity are critical. In this conte xt, we propose a no v el h ybrid approach termed trophallactic optimization algorithm with mark o v random eld renement (T O A-MRF), which syner gistically combines bio-inspired e xploration with probabilistic spatial modeling. The method dra ws inspiration from the trophallaxis mechanism in bee colonies—the process of food e xchange and enzymatic transformation that con v erts collected nectar into ma- ture hone y . This biological analogy translates into a computational frame w ork where virtual bees progressi v ely e xplore the image space from an initial user -dened seed point, e xchange intensity information, and collecti v ely rene re gions of interest through iterati v e transformation mechanisms. The resulting candidate se gmentation is then rened through MRF-based re gularization to produce a spatially coherent and anatomically plausible lesion mask. The proposed T O A-MRF approach of fers se v eral distincti v e adv antages o v er e xisting methods: − Semi-automatic operation requiring only a single user -dened seed point, minimi zing manual interv ention while maintaining clinical control. − Adapti v e e xploration through cooperati v e bee beha vior enabling rob ust se gmentation of heterogeneous re- gions with v arying intensities. − Inte grated classication capability distinguishing between ischemic (old and recent) and hemorrhagic strok e types based on intensity characteristics. − MRF-based spatial re gularization ensuring anatomical continuity and suppressing isol ated f alse detections. Experimental v alidation conducted on clinical brain CT images demons trates the model’ s capabil- ity to ef fecti v ely detect and classify strok e lesions, achie ving high se gmentation accurac y while maintaining computational ef cienc y suitable for clinical deplo yment. The remainder of this paper is or g anized as follo ws: Section 2 re vi e ws related w ork in strok e se g- mentation and bio-inspired optimization. Section 3 presents the detailed method of the T O A-MRF approach. Section 4 descri b e s the e xperimental setup and presents quantitati v e and qualitati v e results. Section 5 discusses the ndings and limitations, follo wed by conclusions and future research directions in section 6. T o conte xtualize T O A–MRF within current medical image analysis trends, recent w ork can be grouped into three complementary directions. First, interacti v e and weakly-supervised se gmentation has g ained trac- tion to reduce annotation costs while k eeping a clinician “in the loop”; scribble- or prompt-dri v en strate gies ha v e demonstrated that sparse user input can guide accurate delineation without full pix el-le v el labels [1]–[4]. Second, as deep models increasingly enter clinical w orko ws, e xplainability and trustw orth y deplo yment ha v e become central research themes, with surv e ys emphasizing limits of salienc y-only e xplanations and the need for go v ernance frame w orks and re gulatory alignment in radiology [5]–[7]. Third, h ybrid pipelines that com- bine learning-based cues with probabilistic renement (e.g., MRF v ariants) continue to be studie d as a w ay to enforce spatial coherence and control f alse positi v es, especially in heterogeneous CT protocols [8], [9]. In par - allel, strok e lesion se gmentation on CT and CT perfusion remains an acti v e topic where recent deep netw orks report strong accurac y b ut still depend on data quality , domain calibrat ion, and careful v alidation across centers [10], [11]. Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 1, April 2026: 131–141 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 133 2. RELA TED W ORK 2.1. Str ok e segmentation techniques Automated strok e lesion se gmentation has been e xtensi v ely studied using di v erse methodological ap- proaches. T raditional methods based on thresholding and re gion gro wing of fer computational ef cienc y b ut struggle with intensity heterogeneity and noise sensiti vity . Acti v e contour models (snak es) and le v el set meth- ods pro vide more sophisticated boundary detection b ut require careful initialization and parameter tuning. Deep learning approaches, particularly con v olutional neural netw orks (CNNs), ha v e achie v ed st ate- of-the-art performance in medical image se gmentation. U-Net architectures [12] and their v ariants ha v e been successfully applied to strok e lesion detection, achi e ving high accurac y on lar ge annotated datasets [13]. Ho w- e v er , these methods require substantial labeled training data, signicant computational resources for training, and may f ace generalization challenges across dif ferent imaging protocols and patient populations. 2.2. Bio-inspir ed optimization in medical imaging Bio-inspired algorithms ha v e found successful applications in v arious medical imaging tasks. The ABC algorithm [14], introduced by Karabog a in 2005, has been applied to image se gmentation, feature selec- tion, and optimization problems [15]. P article sw arm optimization (PSO) [16] has been used for multile v el thresholding and re gistration tasks. Ant colon y optimi zation (A CO) [17] has sho wn promise in edge detection and path planning for sur gical na vig ation. The T O A [18], represents a more recent bio-inspired approach that models the food e xchange and transformation processes in bee colonies. Unlik e standard ABC, T O A empha- sizes the enzymatic transformation aspect of nectar maturation, pro viding a natural mechanism for progressi v e renement of candidate solutions. This characteristic mak es T O A particularly suitable for image se gmentation tasks where gradual re gion renement is desired. 2.3. Mark o v random elds f or spatial r egularization MRF pro vide a principled probabilistic frame w ork for incorporating spatial conte xt in se gmenta- tion [19]. The MRF ener gy minimization approach, typically solv ed through graph cuts [20] or iterated condi- tional modes, balances data delity with spatial smoothness constraints. In medical imaging, MRFs ha v e been e xtensi v ely used for brain tissue classication [21], tumor se gmentation, and multi-modal image fusion. Recent h ybrid approaches combining optimization algorithms with MRF renement ha v e sho wn promising results. Ho we v er , most e xisting methods either focus purely on optimization without spatial mod- eling or apply MRF in isolation without adapti v e e xploration mechanisms. The inte gration of T O A with MRF presented in this w ork addresses this g ap by combining bio-inspired adapti v e e xploration with probabilistic spatial modeling. Be yond algorithmic accurac y , se v eral contrib utions highlight practical considerations for translation to routine care. Interacti v e systems are increasingly accompanied by public implementations and benchmarks that f acilitate reproducibility and rapid prototyping [22], [23]. At the same time, ethical and re gulatory discussions stress requirements for transparenc y , data go v ernance, auditability , and human o v ersight when AI is used as decision support in radiology [24]–[26]. These trends reinforce the rele v ance of interpretable, training-free se gmentation approaches that can complement learning-based models, particularly when data are limited or heterogeneous across acquisition settings. Finally , Ph ysioNet remains a widely used public repository for v alidated ph ysiological and imaging data, supporting reproducible research and e xternal e v aluation [27]. 3. PR OPOSED METHOD The proposed T O A-MRF method inte grates bio-inspired collecti v e intelligence with probabilistic spa- tial modeling to achie v e rob ust semi-automatic strok e l esion se gmentation. The complete processing pipeline consists of six int erconnected stages: image preprocessing, seed-based initialization, bee colon y modeling and e xp l oration, trophallactic e xchange mechanism, MRF-based renement, and automatic classication with visualization. Figure 1 illustrates the o v erall w orko w of the approach. Figure 1 illustrates the o v erall pipeline of the proposed T O A–MRF frame w ork. It sum marizes i) preprocessing steps that enhance CT contrast and suppress noise, ii) interacti v e seed-based initialization, iii) global optimization using the T O A to delineate candidate lesion re gions, and i v) MRF-based renement to enforce spatial coherence and produce the nal lesion mask. The stages of the method and representati v e outputs are further detailed in Figure 2. T r ophallactic optimization algorithm with mark o v r andom eld r enement for str ok e ... (Hayat Berk ok) Evaluation Warning : The document was created with Spire.PDF for Python.
134 ❒ ISSN: 2502-4752 Figure 1. W orko w of the proposed T O A-MRF approach for strok e detection and se gmentation in CT imaging. The pipeline illustrates the progression from preprocessing and seed selection through bee colon y e xploration, enzymatic transformation, hone y reconstruction, and MRF-based renement to the nal se gmentation output 3.1. Image pr epr ocessing The input brain CT scan under goes standardized preprocessing to enhance image quality and prepare data for subsequent analysis: − Grayscale con v ersion and intensity normalization to the range [0, 255]. − Noise reduction through median ltering (k ernel size 3 × 3 ) to suppress imaging artif acts while preserving edge information. − Optional contrast enhancement via adapti v e histogram equalization (CLAHE) to impro v e discrimination between health y and pathological tissue. − Skull stripping or brain e xtraction when necessary to focus analysis on cerebral tissue. The preprocessed image serv es as the e xploration domain for the virtual bee colon y in subsequent stages. 3.2. Seed point selection and initial classication The user initiates the se gmentation process by selecting a single seed point within the suspected lesion re gion. The intensity v al u e at this seed location, denoted I seed , pro vides preliminary classication of strok e type according to empirically established intensity ranges deri v ed from clinical CT imaging characteristics: If 20 < I seed < 70 ⇒ Old ischemic strok e (chronic h ypodense re gion) If 70 < I seed < 100 ⇒ Recent ischemic strok e (acute h ypodense re gion) If I seed > 130 ⇒ Hemorrhagic strok e (h yperdense blood) This seed point serv es as the initial nectar source from which the bee colon y be gins e xploration of neighboring re gions e xhibiting similar intensity characteristics. Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 1, April 2026: 131–141 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 135 3.3. Bee colony modeling and r oles The virtual bee colon y consists of N bees, each represented by a spatial position p i ( x, y ) and associ- ated local intensity v alue I i . Three distinct bee roles collaborate to e xplore and rene the re gion of interest: − Scout bees (30% of population): Perform random e xploration at v arying distances from the seed to disco v er ne w potentially rele v ant re gions. − F orager bees (50% of population): Conduct local e xploitation in the immediate vicinity of the seed, har - v esting pix els with intensity v alues similar to I seed . − Onlook er bees (20% of population): F ollo w high-quality nectar sources identied by scouts and foragers, reinforcing promising re gions through focused e xploration. Each bee e v aluates the quality of its current position using a tness function that combines int ensity similarity with spatial proximity to the seed: f ( p i ) = ω 1 · 1 1 + | I i − I seed | + ω 2 · 1 1 + d ( p i , p seed ) (1) where ω 1 and ω 2 are weighting parameters controlling the relati v e importance of intensity similarity v ersus spatial proximity (typically ω 1 = 0 . 7 , ω 2 = 0 . 3 ), and d ( · , · ) denotes Euclidean distanc e. Higher tness v alues indicate more promising locations for lesion membership. 3.4. T r ophallactic exchange and enzymatic transf ormation The trophall actic mechanism models the biological process of nectar e xchange and enzymatic trans - formation in bee colonies. In each iteration, bees e xchange intensity information with their neighbors, simulat- ing the gradual maturation of nectar into hone y through progressi v e renement: I ( t +1) i = α · I ( t ) i + (1 − α ) · 1 | N ( i ) | X j ∈ N ( i ) I ( t ) j (2) where I ( t ) i represents the intensity v alue at bee position i during iteration t , N ( i ) denotes the neighborhood of bee i (typically 8-connected neighbors), and α is the enzymatic transformation rate ( α = 0 . 6 pro vides balanced renement). This e xchange process ensures gradual homogenization of i ntensity v alues within coherent re gions while maintaining sensiti vity to true boundaries. After con v er gence (typically 50–100 iterations), pix els visited by bees with tness v alues e xceeding a threshold τ ( τ = 0 . 5 ) are mark ed as candidate lesion members, forming the initial “hone y mask” H 0 . 3.5. MRF-based spatial r enement The candidate mask H 0 produced by the bee colon y may contain isolated f alse positi v es or e xhibit boundary irre gularities due to noise and intensity heterogeneity . T o enforce spatial coherence and anatomical plausibility , we apply Mark o v random eld renement through ener gy minimization: E ( M ) = X x D ( x, M x ) + λ X x,y ∈ N V ( M x , M y ) (3) where: − M x ∈ { 0 , 1 } denotes the binary label (lesion/background) at pix el x . − D ( x, M x ) is the data term measuring delity to the initial mask H 0 . − V ( M x , M y ) represents the smoothness term penalizing label discontinuities between neighboring pix els (Potts model). − λ controls the re gularization strength ( λ = 0 . 5 balances delity and smoothness). Ener gy minimization is performed using graph cuts algorithm, producing the nal rened se gmenta- tion mask M ∗ that balances adherence to the bee-generated candidat e re gions with spatial smoothness con- straints aligned to anatomical structure. T r ophallactic optimization algorithm with mark o v r andom eld r enement for str ok e ... (Hayat Berk ok) Evaluation Warning : The document was created with Spire.PDF for Python.
136 ❒ ISSN: 2502-4752 3.6. Classication and visualization The strok e subtype classication is conrmed by computing the mean intensity within the nal mask M ∗ and comparing ag ainst the threshold ranges dened in section 3.2. The se gmentation result is then o v erlaid on the original CT image using color coding for intuiti v e clinical interpretation: − Red: Hemorrhagic strok e. − Dark blue: Old ischemic strok e. − Light purple: Recent ischemic strok e. This color -coded visualization f acilitates rapid assessment by radiologists and supports clinical decision- making in time-critical scenarios. 4. EXPERIMENT AL RESUL TS 4.1. Dataset and experimental setup The e xperimental e v aluation w as conducted using a heterogeneous dataset composed of ischemic and hemorrhagic strok e cases collected from tw o dif ferent sources. Ischemic strok e images were obtai ned from the publicly a v ailable Ph ysioNet database, which pro vides clinically v alidated brain imaging data. Hemorrhagic strok e cases were collected from a neurology clinic and correspond to real clini cal CT scans. The nal dataset consists of 24 patients, including 15 ischemic strok e cases (62.5%) and 9 hemorrhagic strok e cases (37.5%). This combination allo ws the proposed se gmentation method to be e v aluated under realistic and di v erse clinical conditions. All images were anon ymized and used e xclusi v ely for research purposes. All images were normalized to 512 × 512 resolution with 8-bit intensity encoding. The implementa- tion w as de v eloped in Python 3.13, utilizing OpenCV for image processing operations, NumPy for numerical computations, and Matplotlib for visualization. Processing w as performed on a w orkstation with [hardw are specications]. Algorithm parameters were set as follo ws: bee population N = 100 (30% scouts, 50% foragers, 20% onlook ers), maximum iterations=100, tness threshold τ = 0 . 5 , enzymatic rate α = 0 . 6 , MRF re gularization parameter λ = 0 . 5 . These v alues were determined through preliminary e xperiments and maintained constant across all test cases. 4.2. Qualitati v e r esults Figure 2 ill ustrates representati v e results for dif ferent strok e types, sho wing the progression through k e y processing stages. Each ro w presents: (a) the original CT image sho wing the brain scan with visible lesion, (b) the intermediate result after T O A-based bee e xploration (“Miel a v ant MRF”), and (c) the nal se gmentation after MRF renement with color -coded classication o v erlay . V isual ins pection conrms t h a t the T O A stage successfully identies re gions of interest while the MRF renement ef fecti v ely remo v es noise artif acts and smooths boundaries to produce clinically plausible se gmentations. The color coding clearly distinguishes between hemorrhagic strok es (red o v erlay) and ischemic strok es (purple/blue o v erlay), f acilitating rapid clinical interpretation. 4.3. Quantitati v e e v aluation Se gmentation performance w as quantitati v ely assessed using standard metrics computed ag ainst e xpert- annotated ground truth. T able 1 presents a v erage results across strok e subtypes. The results demonstrate con- sistently high performance across both strok e types, with all metrics e xceeding 85%. Hemorrhagic strok es achie v e slightly higher accurac y (Dice=91.2%) due to their higher contrast relati v e to surrounding tissue. The high precision v alues (¿89%) indicate minimal f alse positi v e detections, while strong recall (¿85%) conrms ef fecti v e lesion co v erage. These results v al idate the rob ustness of the T O A-MRF approach e v en wit h minimal user interv ention (single seed point). T able 1. Quantitati v e se gmentation performance by strok e type Strok e T ype Dice (%) IoU (%) Precision (%) Recall (%) Ischemic 87.3 78.5 89.1 85.7 Hemorrhagic 91.2 84.6 92.4 90.1 Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 1, April 2026: 131–141 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 137 (a) (b) (c) Figure 2. T O A-MRF processing stages for strok e se gmentation. F or each case: left panel sho ws the original CT image, center panel sho ws the intermediate mask after T O A processing (“Miel before MRF”), and right panel sho ws the nal MRF-rened se gmentation with color -coded o v erlay: (a) Hemorrhagic strok e with red o v erlay indicating h yperdense blood, (b)-(c) old ischemic strok es with purple o v erlay indicating chronic h ypodense re gions 4.4. Comparati v e analysis T o conte xtualize the performance of T O A-MRF , we compare ag ainst representati v e alternati v e ap- proaches including traditional methods and recent bio-inspired techniques. T able 2 summarizes the comparison on the same test dataset. T able 2. Comparati v e performance analysis Method A vg Dice (%) A vg IoU (%) User Input T ime (s) Otsu thresholding 71.4 60.2 None 0.2 Re gion gro wing 78.6 68.3 1 seed 1.5 FCM + ABC 83.7 74.1 None 8.3 U-Net (CNN) 92.8 86.4 None 0.8 T O A-MRF (Proposed) 89.3 81.6 1 seed 4.2 The comparison re v eals that T O A-MRF achie v es competiti v e accurac y with deep learning approaches (Dice: 89.3% vs. 92.8% for U-Net) while maintaining the adv antages of requiring minimal user input (single seed) and no training phase. T raditional methods sho w lo wer performance due to limited ability to handle comple x intensity distrib utions. The computational time of 4.2 seconds represents a practical balance between accurac y and ef cienc y for clinical application. 5. DISCUSSION The e xperimental results demonstrate that the T O A-MRF approach successfully combines bio-inspired adapti v e e xploration with probabilistic spatial re gularization to achie v e rob ust strok e lesion se gmentation. Se v- eral k e y ndings merit detailed discussion. T r ophallactic optimization algorithm with mark o v r andom eld r enement for str ok e ... (Hayat Berk ok) Evaluation Warning : The document was created with Spire.PDF for Python.
138 ❒ ISSN: 2502-4752 Wh y hemorrhagic lesions are se gmented more accurately: As reected by the quantitati v e and qual- itati v e results, hemorrhagic strok es tend to achie v e higher Dice/IoU than ischemic strok es. This beha vior is consistent with the ph ysics of non-contrast CT : acut e hemorrhage appears h yperdense relati v e to surround- ing parench yma, creating sharper intensity discontinuities that are easier to capture with both re gion-based optimization and edge-consistenc y priors. In contrast, ischemic lesions are often h ypodense with subtle bound- aries and can o v erlap in intensity with cerebrospinal uid, v entricles, or chronic white-matter changes, which increases the ambiguity of intensity-thresholding and may lead to partial under -s e gm entation. The MRF rene- ment mitig ates isolated f alse positi v es by enforcing local spatial coherence, b ut residual uncertainty remains when lesion-to-background contrast is lo w , moti v ating adapti v e calibration strate gies and multi-feature cues for future w ork [8], [9]. Seed sensiti vity and user interaction: T O A–MRF is intentionally designed as a lightweight, training- free method where a clinician pro vides a small number of seeds to indicate lesion and background. While this interaction impro v es controllability and can reduce the need for lar ge annotated datasets, it also implies sensiti vity to seed placement, especially in ischemic cases where l esion borders are weak. This limitation is well recognized in interacti v e se gmentation literature, where rob ustness is impro v ed by strate gies such as multi-seed initialization, uncertainty-guided renement, and incorporating weak supervision such as scribbles or prompts [1], [2]. In practice, a simple mitig ation is to allo w multiple fore ground seeds distrib uted across the suspected lesion re gion and to pro vide immediate visual feedback; this aligns with radiology w orko ws where rapid corrections are preferable to length y training pipelines. Future w ork may further automate seed suggestion using lightweight heuristics (e.g., intensity outlier detection) or h ybrid prompting mechani sms, while k eeping the core algorithm transparent and easy to deplo y . Generalization across scanners and x ed thresholds: A k e y design choice in this w ork is the use of x ed intensity ranges for coarse s trok e-type characterization. Although ef fecti v e in the studied dataset, such thresholds may shift with scanner v endor , reconstruction k ernel, slice thickness, or calibration dif ferences, which can reduce generalization across centers. This limitation moti v ates tw o complementary directions: i) data-dri v en normalization or histogram standardization prior to thresholding, and ii) adapti v e threshold selec- tion using image-specic statistics or learned calibration maps. Recent CT strok e se gmentation studies empha- size that cross-domain rob ustness and e xternal v alidation are critical for deplo yment, and that e v en strong deep models can de grade under protocol shift without proper calibration [10], [11]. Importantly , the training-free nature of T O A–MRF f acilitates rapid recalibration because on l y a small set of parameters and rules must be adjusted rather than retraining a netw ork. Computational ef cienc y and scalability to 3D v olumes: The reported runtime (approximately a fe w seconds per 2D slice on standard hardw are) is compatible with interacti v e use. F or full 3D CT v olumes, to- tal processing time will scale with the number of slices; ho we v er , the algorithm is amenable to acceleration through parallel processing because slice-le v el optimization and MRF renement can be computed indepen- dently before optional 3D post-processing. A practical deplo yment strate gy is there fore to run T O A–MRF slice-by-slice with lightweight temporal/3D smoothing, or to restrict processing to a radiologist-dened re- gion of interest. These strate gies preserv e the interacti v e nature of the approach while enabling near real-time v olumetric assessment. Clinical inte gration, interpretability , and go v ernance: From a clinical perspecti v e, T O A–MRF can be vie wed as a decision-support tool that complements the radiologist rather than replacing e xpert judgment. Its seed-based interaction naturally ts into routine C T reading, where clinicians can quickly indicate a suspected lesion and obtain an interpretable mask that is rened by e xplicit spatial constraints. Unlik e black-box con- v olutional netw orks, the method pro vides transparent intermediate stages (preprocessing, optimization-dri v en re gion formation, probabilistic renement), which can f acilitate user trust and error auditing. This is aligned with recent guidance emphasizing e xplainability , human o v ersight, and go v ernance frame w orks for safe clini- cal adoption of AI in radiology [5]–[7]. In lo w-resource settings where lar ge annotated datasets and high-end GPUs may be una v ailable, training-free approaches remain attracti v e because the y can be deplo yed with mini- mal infrastructure while still of fering clinically useful se gmentation assistance. Limitations and future directions: Despite promising results, this study has se v eral limitations. First, although the dataset combines public ischemic cases with real clinical hemorrhagic scans, the total cohort size remains modest, and broader multi-center v alidation is required to c o n rm rob ustness across di v erse popu- lations and CT protocols. Second, the method currently depends on user -pro vided seeds and x ed intensity heuristics; inte grating adapti v e calibration, multi- seed strate gies, and uncertainty feedback w ould reduce sen- Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 1, April 2026: 131–141 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 ❒ 139 siti vity and impro v e reliability . Finally , future e v aluation should include statis tical signicance testing ag ainst strong baselines and e xtended v olumetric e xperiments, as recommended in recent se gmentation benchmarking and clinical translation studies [1]–[10]. 5.1. Str engths of the appr oach − Semi-automatic operation with minimal user b urden: The single-seed init ialization represents a signicant practical adv antage o v er fully manual se gmentation or methods requiring e xtensi v e parameter tuning. This reduces radiologist w orkload while maintaining clinical control. − Biological plausibility and interpretability: The trophallactic mechanis m pro vides an intuiti v e analogy to natural collecti v e beha vior , making the algorithm’ s operation more interpreta b l e than black-box deep learn- ing models. This transparenc y is v aluable in clinical settings where e xplainability is increasingly important. − Syner gistic combination of T O A and MRF: The T O A component e f ciently e xplores heterogeneous inten- sity re gions while the MRF renement impose s anatomical coherence. This tw o-stage approach le v erages complementary strengths: adapti v e e xploration follo wed by principled spatial modeling. − Inte grated classication capability: The intensity-based strok e type classication enables automatic dif fer - entiation between ischemic and hemorrhagic cases wi thout requiring separate supervised class iers. This simplies the clinical w orko w . − No training requirement: Unlik e deep learning approaches that demand lar ge annotated datasets and com- putational resources for training, T O A-MRF operates directly on ne w images with x ed parameters. This f acilitates deplo yment in resource-limited settings and across dif ferent imaging protocols. 5.2. Limitations and challenges − Seed placement sensiti vity: The method’ s performance depends on appropriate seed selection within the lesion. Poorly positioned seeds (e.g., on boundaries or in artif act re gions) can lea d to suboptimal se gmen- tation. Future w ork could e xplore automatic seed detection or multi-seed strate gies. − Fix ed intensity thresholds: The empirical ranges for strok e classication may not generalize across all CT scanners and imaging prot o c ols. Adapti v e threshold learning or normalization strate gies could impro v e rob ustness. − Computational ef cienc y for lar ge images: The iterati v e bee e xploration and trophallactic e xchange pro- cesses scale with image size and bee population. F or v ery lar ge 3D v olumes, computational time could become prohibiti v e without optimization (e.g., GPU acceleration, hierarchical processing). − Limited v alidation dataset: The current e v aluation w as conducted on a single-cent er dataset. Multi-center v alidation with di v erse imaging protocols and patient populations w ould s trengthen generalizability claims. − Small lesion detection: V ery small lesions (fe w pix els) may be missed or incorrectly classied due to limited spatial information. Multi-scale processing or attention mechanisms could address this limitation. 5.3. Futur e r esear ch dir ections − Extension to 3D v olumetric analysis with slice-by-slice or full 3D bee colon y e xploration. − Inte gration with deep learning features: using CNN-e xtracted features within the T O A tness function could combine model-free e xploration with learned representations. − Adapti v e parameter learning through reinforcement learning or meta-learning to automatically tune algo- rithm parameters for dif ferent imaging conditions. − Clinical v alidation studies with radiologist e v aluation of se gmentation quality and time sa vings in real clinical w orko ws. − Application to other neuroimaging tasks such as tumor se gmentation, white matter lesion detection, or v ascular structure e xtraction. Ov erall, the T O A-MRF approach represents a promising direction for semi-automatic medical image se gmentation that balances accurac y , interpretability , and practical usability . While not yet matching the peak performance of state-of-t he-art deep learning models, it of fers distinct adv antages in scenarios where training data is limited, computational resources are constrained, or model transparenc y is prioritized. T r ophallactic optimization algorithm with mark o v r andom eld r enement for str ok e ... (Hayat Berk ok) Evaluation Warning : The document was created with Spire.PDF for Python.
140 ❒ ISSN: 2502-4752 6. CONCLUSION This paper introduced T O A-MRF , a no v el h ybrid approach for semi-automatic strok e lesion se gmenta- tion in brain CT images that syner gistically combines the T O A with Mark o v random eld spatial re gularization. The method dra ws inspiration from the collecti v e nectar e xchange and transformation beha vior of bee colonies, translating this biological mechanism into a computational frame w ork for progressi v e image e xploration and renement. The proposed approach addresses k e y challenges in medical image se gmentation by requiring only minimal user input (a single seed point) while achie ving competiti v e accurac y through adapti v e bio-inspired e xploration follo wed by probabilistic spatial modeling. Experimental v alidation on clinical CT data demon- strated Dice coef cients of 87.3% for ischemic strok es and 91.2% for hemorrhagic strok es, with o v er all perfor - mance approaching that of supervised deep learning methods while maintaining adv antages in interpretability , training-free operation, and clinical usability . The inte gration of T O A ’ s collecti v e intelligence with MRF’ s spa- tial re gularization pro v ed ef fecti v e in handling the heterogeneous intensity distrib utions and anatomical v ari- ability characteristic of cerebro v ascular lesions. The method successfully balances e xploration and e xploitation through role-based bee cooperation, progressi v ely renes re gions through trophallactic e xchange, and enforces anatomical plausibility through ener gy-based spatial smoothing. While certain limitations remain—particularly re g arding seed placement sensiti vity and general iza- tion across di v ers e imaging protocols—the T O A-MRF frame w ork establishes a foundation for future research in bio-inspired medical image analysis. Potential e xtensions include 3D v olumetric processing, inte gration with deep learning features, adapti v e parameter opt imization, and application to other neuroimaging se gmen- tation tasks. In conclusion, the T O A-MRF approach of fers a practical and theoretically grounded solution for computer -aided strok e diagnosis, contrib uting to the gro wing body of w ork at the intersection of bio-inspired optimization and medical imaging. By reducing manual se gmentation b urden while maintaining clinical control and interpretability , such methods ha v e the potential to support radiologists in time-critical diagnostic scenarios and impro v e patient outcomes through f aster , more accurate strok e assessment. REFERENCES [1] Y . Qu, T . Lu, S. Zhang, and G. W ang, “ScribSD+: Scribble-supervised medical image se gmentation based on simultaneous multi- scale kno wledge distillation and class-wise contrasti v e re gularization, ” Computerized Medical Ima ging and Gr aphics , 2024, doi: 10.1016/j.compmedimag.2024.102236. [2] H. E. W ong, M. Rakic, J. Guttag, and A. V . Dalca, “ScribblePrompt: F ast and Fle xible Interacti v e Se gmentation for An y Biomedical Image, ” in Pr oc. ECCV , 2024. [Online]. A v ailable: https://www .ecv a.net/papers/eccv 2024/papers ECCV/papers/05664.pdf. [3] T . W ang et al. , “ScribbleVS: scribble-supervised medical image se gmentation via re gional pseudo la bels dif fusion and dynamic competiti v e selection, ” arXiv pr eprint arXiv:2411.10237 , 2024. [Online]. A v ailable: https://arxi v .or g/abs/2411.10237. [4] X. Luo et al. , “Scribble-supervised medical image se gmentation via dual-branch netw ork, ” arXiv pr eprint arXiv:2203.02106 , 2022. [Online]. A v ailable: https://arxi v .or g/abs/2203.02106. [5] K. Borys et al. , “Explainable AI in medical imaging: an o v ervie w for clinical practitioners, ” Eur opean J ournal of Radiolo gy , v ol. 162, Art. 110786, 2023, doi:10.1016/j.ejrad.2023.110786. [6] F . M. Aldhafeeri “Go v erning articial intelligence in radiology: a systematic re vie w of ethical, le g al, and re gulatory frame w orks, ” Dia gnostics , v ol. 15, no. 18, Art. 2300, 2025, doi: 10.3390/diagnostics15182300. [7] E. H. Houssein et al. , “Explainable articial intelligence for me dical imaging systems using deep learning methods: A comprehen- si v e re vie w , ” Cluster Computing , 2025, doi: 10.1007/s10586-025-05281-5. [8] H. W ang and L. Ma, “MCMC algorithm based on Mark o v random eld in image se gmentation, ” PLOS ONE , v ol. 18, no. 12, e0296031, 2023, doi: 10.1371/journal.pone.0296031. [9] A. Daoudi and S. Mahmoudi, “Enhancing brain se gmentation in MRI through inte gration of hidden mark o v random eld and whale optimization algorithm, ” Computer s , v ol. 13, no. 5, Art. 124, 2024, doi: 10.3390/computers13050124. [10] A. Kandpal, R. K. Gupta, and A. Singh, “Ischemic strok e lesion se gmentation on multiparametric CT perfusion maps using deep neural netw ork, ” AI , v ol. 6, no. 1, Art. 15, 2025, doi: 10.3390/ai6010015. [11] O. I. Alirr , “Ischemic strok e lesion core se gmentation from CT perfusion scans using atte ntion ResUnet deep learning, ” J ournal of Ima ging Informatics in Medicine , v ol. 38, pp. 3507–3516, 2025, doi: 10.1007/s10278-025-01407-8. [12] O. Ronneber ger , P . Fis cher , and T . Brox, “U-net: Con v olutiona l netw orks for biomedical image se gmentation, ” in Medical Ima g e Computing and Computer -Assisted Intervention (MICCAI) , LNCS 9351, pp. 234–241, 2015. [13] L. Chen, P . Bentle y , and D. Rueck ert, “Fully automatic acute ischemic lesion se gmentation in D WI using con v olutional neural netw orks, ” Neur oIma g e: Clinical , v ol. 15, pp. 633–643, 2017. [14] D. Karabog a, “ An idea based on hone y bee sw arm for numerical optimization, ” T echnical Report TR06, Erciyes Uni v ersity , Engi- neering F aculty , Computer Engineering Department, 2005. [15] M. T aherdangk oo, M. Y azdi and M. H. Rezv ani, ”Se gmentation of MR brain images using FCM impro v ed by articial bee colon y (ABC) algorithm, ” Proceedings of the 10th IEEE International Conference on Information T echnology and Applications in Biomedicine, Corfu, Greece, 2010, pp. 1-5, doi: 10.1109/IT AB.2010.5687803. Indonesian J Elec Eng & Comp Sci, V ol. 42, No. 1, April 2026: 131–141 Evaluation Warning : The document was created with Spire.PDF for Python.