PathMNIST-XAI: Explainable CNN for Medical Imaging

A lightweight yet high-performing CNN model designed for the PathMNIST medical imaging dataset.
This model integrates explainable AI (XAI) techniques using Integrated Gradients for transparent decision-making, while also leveraging SQLite for scalable attribution storage.

With over 91% test accuracy and 97% training accuracy, the model demonstrates state-of-the-art efficiency, scalability, and explainability in pathology image classification.

---

🔍 Introduction & Overview

Medical image classification is critical for computer-aided diagnosis (CAD). However, black-box models often lack interpretability, limiting trust in clinical adoption.

This project introduces an improved CNN architecture with explainability baked in:

• Trained on the PathMNIST dataset (107,180 images).
• Achieved 91.14% test accuracy in just 25 epochs.
• Uses Captum’s Integrated Gradients for interpretable attributions.
• Stores explanations in a scalable SQLite database for downstream analysis.

👉 Keywords: AI Model, Deep Learning, CNN, Scalable AI, Lightweight Model, Medical Imaging, Explainable AI, Pathology, Integrated Gradients, SQLite

---

⚡ Key Features

• ✅ High Accuracy: 97.11% training, 91.14% test accuracy.
• ✅ Lightweight & Scalable: Efficient CNN architecture deployable on edge and HPC systems.
• ✅ Explainability Built-in: Integrated Gradients for transparent predictions.
• ✅ Data Augmentation: Horizontal flips & rotations improve generalization.
• ✅ Scalable Attribution Storage: Explanations stored in SQLite for reproducibility.
• ✅ Early Stopping + Scheduler: Prevents overfitting, ensures stable training.

---

📈 Results & Benchmarks

• Final Test Accuracy: 91.14%
• Peak Training Accuracy: 97.11%
• Outperforms baseline CNNs for PathMNIST while maintaining lightweight scalability.

Metric	Our Model
Training Accuracy	97.11%
Test Accuracy	91.14%
Epochs	25
Explainability	✅ IG Maps
DB Storage Support	✅ SQLite

---

🏗️ Model Architecture

The architecture includes:

• Convolutional blocks with BatchNorm + ReLU for stability.
• Dropout layers for regularization.
• Fully connected layers for robust classification into 9 classes.

---

📊 Training Performance

Training ran for 25 epochs with augmentation, Adam optimizer, and ReduceLROnPlateau scheduler.

• Accuracy improved from 77.03% → 97.11% during training.
• Loss decreased consistently, showing strong convergence.

---

⚙️ Installation & Usage

1️⃣ Clone Repository

git clone https://github.com/ShreyaVijaykumar/PathMNIST-XAI-Lightweight-Explainable-CNN-for-Medical-Imaging
cd pathmnist-explainable-cnn

2️⃣ Install Requirements

pip install -r requirements.txt

3️⃣ Run Training

python train.py

4️⃣ Run Evaluation + Explanations

python evaluate.py

---

🔑 Keywords

#AI #DeepLearning #ExplainableAI #XAI #CNN #MedicalAI #Pathology #MedicalImaging
#PyTorch #PathMNIST #LightweightAI #ScalableAI #EdgeAI #OpenSourceAI #MLforHealth
#IntegratedGradients #Captum #DatabaseAI #SQLite #HealthcareAI #ComputerVision