1. Introduction

Medical image segmentation plays a crucial role in advancing healthcare technologies. By segmenting anatomical structures or pathological regions, clinicians significantly enhance diagnostic accuracy, improve treatment planning, and monitor disease progression with greater precision. Although deep neural networks (DNNs) such as U-Net and fully convolutional neural networks (FCNN) deliver high accuracy in segmentation tasks, they often lack explainability, which is a major barrier to clinical trust and adoption.

Since clinicians frequently hesitate to trust outputs generated by black-box AI models, the need for interpretable systems has become increasingly urgent. For this reason, healthcare applications demand models that offer both high performance and clear reasoning. ExShall-CNN is designed to address this exact gap. By bridging the power of shallow convolutional neural networks (CNNs) with state-of-the-art interpretable machine learning techniques, ExShall-CNN delivers robust performance and achieves transparent, clinically meaningful results.

2. The Necessity of Explainability in Medical Image Segmentation

Medical imaging serves as a cornerstone in healthcare decision-making. Therefore, AI models utilized for these tasks must not only be accurate but also demonstrate interpretability. Explainable AI is particularly crucial, as it directly influences trust, regulatory compliance, and real-world applicability.

Challenges with Deep Black-Box Models

Deep learning models like U-Net excel in segmentation, yet they fail to offer insights into their decision-making processes. This creates significant issues:

• Clinicians find it difficult to validate or verify model predictions due to a lack of transparency.
• Regulatory agencies hesitate to approve opaque systems because accountability remains elusive.
• Adoption becomes constrained in high-stakes applications, such as oncology and surgical planning, where understanding the reasoning behind a decision is paramount.

Explainable Models: Bridging the Gap

Explainable AI bridges these challenges by aligning outputs with interpretable features. Therefore, models such as ExShall-CNN successfully build trust among healthcare professionals, which, in turn, facilitates smoother adoption into clinical workflows. Furthermore, its transparent design not only allows clinicians to validate AI-generated outputs with confidence but also ensures that the results remain closely aligned with observable medical patterns, enhancing overall usability.

3. Methodology of ExShall-CNN for Medical Image Segmentation

ExShall-CNN leverages an innovative methodological framework that combines interpretability, efficiency, and accuracy. By focusing on modular architecture, specialized feature mapping, and dataset diversity, the model achieves its unique balance of transparency and performance.

Modular Architecture for Interpretability in Medical Image Segmentation

The architecture of ExShall-CNN consists of three primary module types, each playing a crucial role:

1. Conv Modules: These foundational modules focus on linear operations like summation and subtraction to extract raw features from image pixels. Because these operations are simple, they are inherently interpretable.
2. Log-Conv-Exp (LCE) Modules: These modules expand on linear features, incorporating complex transformations such as multiplication, division, and local feature extraction. Consequently, they enhance segmentation precision while maintaining transparency.
3. Conv-Log-Conv-Exp (CLCE) Modules: These advanced modules emphasize spatial and statistical relationships. By leveraging kernels such as Chi-Squared and Radial Basis Function (RBF), CLCE modules refine intricate segmentation boundaries effectively.

Kernel Design and Efficiency in Medical Image Segmentation

To ensure sufficient receptive fields while maintaining a shallow design, ExShall-CNN employs larger kernel sizes (ranging from 1 to 25). This approach compensates for the lack of depth without sacrificing spatial coverage, enabling accurate segmentation even in complex scenarios.

Feature Mapping with Transparency

ExShall-CNN excels in feature mapping by extracting both global and local patterns from pixel values and their arrangements. As a result, the model produces intermediate outputs that can be directly visualized and understood by clinicians. This capability makes it easier for healthcare professionals to relate segmentation outputs to real-world medical patterns.

Diverse Datasets for Robust Evaluation

ExShall-CNN has been rigorously tested on two benchmark datasets:

1. Retina Blood Vessel Dataset: This dataset focuses on vascular segmentation, with 100 annotated retinal fundus images. The dataset is essential for diagnosing retinal conditions like diabetic retinopathy.
2. ISIC Dataset: Comprising over 7,700 skin lesion images, the ISIC dataset is invaluable for melanoma detection. For this study, ExShall-CNN was tested on a curated subset of 100 images, ensuring robust evaluation.

4. Performance Evaluation in Medical Image Segmentation

The performance of ExShall-CNN was assessed against top-performing models like FCNN and U-Net. The evaluation utilized widely accepted metrics, including Jaccard and Sørensen-Dice coefficients, to quantify segmentation accuracy and reliability.

Table: Retina Blood Vessel Dataset Performance

Model	Parameters	Training Jaccard (%)	Testing Jaccard (%)	Testing Dice (%)
FCNN	54,304,086	63.3	56.8	72.4
U-Net	31,043,521	63.5	65.9	79.0
ExShall-CNN	39,698	56.2	58.3	73.6

Generalization Capability

ExShall-CNN demonstrated superior generalization on unseen datasets compared to FCNN. This achievement can be attributed to its smaller parameter space, which significantly reduces the risk of overfitting, ensuring that the model remains reliable in diverse clinical applications.

5. Explainability: A Defining Strength of ExShall-CNN

ExShall-CNN goes beyond accuracy by emphasizing interpretability, which is its defining strength. The model employs direct visualization techniques to illustrate the influence of specific modules, making its reasoning comprehensible to clinicians.

Impactful Modules

A scoring system identified modules with the highest contributions to segmentation accuracy. These modules, such as Conv-1 and LCE-9, were visualized to demonstrate their transformations and explain their significance.

Table: Most Impactful Modules

Module Index	Type	Kernel Size	Impact Score
1	Conv	1	0.56
7	Conv	21	0.46
9	LCE	1	0.50
15	LCE	17	0.46
19	CLCE	3	0.40

Visual Outputs for Transparency

Through visual representations, ExShall-CNN allows clinicians to see exactly how it reaches its conclusions. These outputs not only clarify model reasoning but also build trust by aligning results with observable medical features.

6. Benefits and Limitations

Advantages

• Transparency: ExShall-CNN’s visual outputs enable clinicians to understand and trust its decisions.
• Generalization: The model performs well on unseen datasets, reducing overfitting risks.
• Clinical Relevance: Segmentation results align with real-world patterns, ensuring usability in medical settings.

Limitations

• Limited scalability for larger datasets.
• Dependence on hand-crafted features may constrain flexibility in dynamic environments.

7. Future Directions

Future research aims to enhance ExShall-CNN’s scalability and expand its capabilities for multi-resolution image analysis. By testing the model on broader datasets, researchers plan to refine its performance and further validate its applicability across diverse medical domains.

8. Conclusion

ExShall-CNN sets a new benchmark for medical image segmentation by blending accuracy with interpretability. While deep models like U-Net achieve slightly higher raw accuracy, ExShall-CNN’s transparent architecture ensures better generalization and clinical usability. Its innovative design and modular approach make it a valuable tool for healthcare professionals seeking reliable and explainable AI solutions.

Click here to see more blogs.

Reference

Khalkhali, V., Azim, S.M., Dehzangi, I. ExShall-CNN: An Explainable Shallow Convolutional Neural Network for Medical Image Segmentation. Machine Learning and Knowledge Extraction. 2025, 7, 19. DOI: 10.3390/make7010019. License: CC BY 4.0.