Applications of Explainable AI in Disease Diagnosis

This research initiative explores the integration of Deep Learning with Explainable Artificial Intelligence (XAI) to build transparent, interpretable, and trustworthy diagnostic tools in medical imaging. By leveraging Grad-CAM with multiple convolutional neural network architectures, the project provides visual insights into the decision-making process of AI models.

📚 This work resulted in two published research papers focusing on interpretability and performance of segmentation models in disease diagnosis.

---

🎯 Objective

To develop robust medical image segmentation models and integrate Grad-CAM to visualize model focus areas—bridging the gap between computational efficiency and clinical interpretability.

---

🧬 Deep Learning Architectures Used

• U-Net — Classic semantic segmentation architecture
• MultiResU-Net — Incorporates multi-resolution pathways for fine detail capture
• DCU-Net — Utilizes dense connections for feature reuse and efficiency
• VU-Net (Proposed) — Hybrid of VNet and U-Net tailored for multiclass segmentation

---

🗂 Datasets

• 🫀 CT Heart Disease Dataset
• 🧫 Breast Ultrasound Images Dataset
• 🧬 LiTS17 Liver Tumor Segmentation Challenge Dataset

Loss Functions Used:

• Binary Cross Entropy
• Dice Loss
• Binary Focal Loss

---

🔍 Explainability via Grad-CAM

Gradient-weighted Class Activation Mapping (Grad-CAM) was integrated to generate heatmaps identifying key decision areas within medical images. These heatmaps:

• Validate the model’s reasoning against clinical understanding
• Aid in debugging and training improvements
• Enhance transparency and clinical adoption of AI tools

---

📊 Key Findings

• MultiResU-Net consistently delivered high segmentation accuracy
• VU-Net, the proposed model, performed best on multiclass tasks (notably LiTS17)
• Grad-CAM visualizations confirmed medical relevance in model attention maps

---

🧰 Tech Stack

• Languages: Python
• Libraries: TensorFlow, Keras, NumPy, OpenCV, Matplotlib
• Environment: Google Colab
• Formats: PNG, NII

---

📄 Publications

1. Applications of Explainable AI in Disease Diagnosis using CNNs and Grad-CAM
   Presented at: 2023 IEEE Symposium on Computational Intelligence in Medicine and Healthcare (CIMED)
   DOI: 10.1109/10725563

2. Employing Grad-CAM in DL Models for Tumor Segmentation and Visual Explanation: An Empirical Study
   In: Data Science and Applications, Springer, 2024
   DOI: 10.1007/978-981-96-1188-1_35

---

👨‍💻 Contributors

• Pranjal Singh Katiyar 🔗Linkedin
• Ramakrishnananda 🔗Linkedin
• Supervisor: Dr. Rosy Sarmah, 🔗Associate Professor, Dept. of CSE, Tezpur University

---

🚀 Future Directions

• Extend VU-Net for 3D volumetric segmentation
• Explore transformer-based architectures (e.g., TransUNet)
• Integrate real-time clinician feedback for validation
• Expand to cross-modality medical imaging (X-ray, MRI)