🧠 Deep Learning for Tumor Classification

Medical Image Analysis for Benign and Malignant Tumor Classification using PyTorch and the MedMNIST dataset.

📌 Project Overview

This project leverages deep learning to classify Breast MRI images as either benign or malignant tumors. The model was built using PyTorch and trained on the BreastMNIST subset of the MedMNIST dataset. The goal is to provide a lightweight, interpretable, and efficient pipeline for medical image classification tasks.

🛠️ Tools & Technologies

The project uses the following core tools:

• PyTorch: Primary framework for defining and training the deep learning model.
• Torchvision.transforms: Used for data preprocessing and augmentation (e.g., normalization, resizing).
• TQDM: For visually appealing progress bars in training loops.
• NumPy: For mathematical array/tensor manipulation and helper operations.
• MedMNIST: A curated collection of lightweight medical image datasets. We use the BreastMNIST subset.
• medmnist.INFO & Evaluator: For dataset metadata and built-in evaluation with metrics like accuracy and AUC.

📂 Dataset

We used the BreastMNIST dataset from MedMNIST, which contains grayscale breast MRI images labeled as:

• 0 = Benign tumor
• 1 = Malignant tumor

Each image is standardized and simplified to 28×28 pixels, enabling quick experimentation and benchmarking.

MedMNIST is designed for educational and research purposes. The BreastMNIST subset is derived from the Breast Cancer MRI Cohort (TCIA).

🧠 Model Architecture

A simple but effective Convolutional Neural Network (CNN) was built in PyTorch with the following layers:

• Convolution + ReLU + MaxPooling
• Fully Connected (FC) layers
• Sigmoid output for binary classification

Training pipeline includes:

• Binary Cross-Entropy Loss
• Adam optimizer
• Accuracy tracking and MedMNIST's built-in evaluator

🚀 How to Run

1. Clone the repository:

   git clone https://github.com/yourusername/tumor-classification-pytorch.git
   cd tumor-classification-pytorch
   

2. Install dependencies:

   pip install torch torchvision medmnist tqdm numpy

3. Run training:

   python train.py

4. Run evaluation:

   python evaluate.py

📈 Results

• Achieved 86% validation accuracy on binary classification of breast tumor MRI scans.
• Trained using only a few convolutional layers with minimal preprocessing, demonstrating strong performance on a lightweight dataset.

🔍 Sample Code Snippet

from medmnist import INFO, Evaluator
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
from tqdm import tqdm

🧪 Future Work

• Increase accuracy by:
  • 📉 Using a smarter loss function: Implemented BCEWithLogitsLoss with pos_weight to handle class imbalance more effectively and reduce false negatives.
  • 🕒 Training longer: Increased the number of training epochs (e.g., 25–50) and used early stopping based on validation loss to avoid overfitting.
  • 🧪 Hyperparameter tuning:
    • Learning rates: 1e-3, 1e-4, 1e-5
    • Batch sizes: 32, 64
    • Optimizers: Adam, AdamW, RMSprop
    • Learning rate scheduler:

      scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.5)

• 🔍 Integrate Grad-CAM or other explainability tools for visual model interpretability
• 🧠 Experiment with deeper CNNs or pre-trained architectures like ResNet or EfficientNet
• 🚀 Deploy as an interactive web app using Streamlit or Flask

Thanks for Checking this out!

Disclaimer: This project is for educational and research purposes only and is not intended for clinical use.