Prash2410/cancerclassification_customcnn

---

Lung Cancer Classification using a Custom Residual CNN (PyTorch)

Project Overview

This project presents a deep learning framework for multi-class lung cancer classification from histopathological images using a custom Residual Convolutional Neural Network (ResCNN) implemented entirely in PyTorch.

The model distinguishes between healthy lung tissue and two major lung cancer subtypes while leveraging ResNet-inspired residual learning to improve optimization stability, gradient propagation, and convergence speed. The objective is to establish a lightweight yet effective baseline for automated lung cancer diagnosis from histopathological data.

---

Problem Statement

Lung cancer remains one of the leading causes of cancer-related mortality worldwide. Histopathological examination serves as the gold standard for diagnosis; however, manual analysis is both time-intensive and highly dependent on expert interpretation.

This project aims to develop an automated deep learning system capable of classifying histopathological lung tissue images into multiple categories, thereby supporting computer-aided diagnosis and reducing diagnostic workload.

---

Dataset

The dataset consists of three classes:

Class Label	Description
lung_n	Normal Lung Tissue
lung_aca	Lung Adenocarcinoma
lung_scc	Lung Squamous Cell Carcinoma

---

Data Preparation & Training Pipeline

Data Preparation

The following preprocessing workflow was adopted:

• Organized the dataset into training and validation splits.
• Applied image normalization using PyTorch transforms.
• Built efficient DataLoader pipelines for batched processing.
• Preserved class integrity throughout experimentation.
• Maintained reproducibility through consistent preprocessing procedures.

Model Development

A custom residual architecture was designed with the following structure:

• Progressive feature extraction: 3 → 8 → 16 → 32 → 64 → 128 channels.
• Residual learning through skip connections.
• Projection shortcuts (1×1 convolutions) whenever feature dimensions changed.
• Batch Normalization after every convolutional layer.
• Global Average Pooling for parameter efficiency.
• Fully-connected classifier with dropout regularization.

Training Configuration

Component	Configuration
Framework	PyTorch
Optimizer	Adam
Initial Learning Rate	1e-4
Scheduler	ReduceLROnPlateau
Loss Function	CrossEntropyLoss
Dropout	0.2
Epochs	20

The learning rate schedule evolved as follows:

1e-4 → 5e-5 → 2.5e-5 → 1.25e-5 → 6.3e-6 → 3.1e-6

This adaptive strategy enabled smoother convergence during later stages of training.

---

Model Architecture

The proposed network consists of five residual stages followed by a lightweight classification head.

Input Image
      ↓
Residual Block (3 → 8)
MaxPool(2)

Residual Block (8 → 16)
MaxPool(2)

Residual Block (16 → 32)
MaxPool(2)

Residual Block (32 → 64)
MaxPool(2)

Residual Block (64 → 128)
MaxPool(2)

Adaptive Average Pooling (1×1)

Flatten

Linear (128 → 256)
BatchNorm
ReLU
Dropout (0.2)

Linear (256 → 128)
BatchNorm
ReLU
Dropout (0.2)

Linear (128 → 3)

---

Residual Learning

The model incorporates ResNet-inspired residual blocks:

Output = F(x) + x

where:

• F(x) denotes the learned residual mapping.
• x denotes the shortcut connection.

Whenever channel dimensions change, projection shortcuts using 1×1 convolutions and Batch Normalization are employed.

Benefits of Residual Connections

• Faster convergence during optimization.
• Improved gradient flow across deeper layers.
• Reduced vanishing-gradient effects.
• Easier training of deeper architectures.
• Better feature reuse and representation learning.
• Increased training stability.

The inclusion of skip connections played a significant role in achieving strong performance while maintaining a lightweight model footprint.

---

Results

Best Performance Metrics

Metric	Value
Maximum Validation Accuracy	96.46%
Best Validation Loss	0.1268
Best Epoch	12
Overall Classification Accuracy	95.00%
Macro F1-Score	0.95
Weighted F1-Score	0.95

---

Training Performance

The network demonstrated rapid and stable convergence, exceeding 95% validation accuracy while maintaining low validation loss.

The best-performing checkpoint was obtained at:

Epoch               : 12
Validation Accuracy : 96.46%
Validation Loss     : 0.1268

Training Curves

Training Loss

Figure 1: Training loss across epochs.

---

Testing Loss

Figure 2: Testing loss across epochs.

---

Testing Accuracy

Figure 3: Test accuracy across epochs.

The combination of:

• Residual learning,
• Batch Normalization,
• Adaptive learning-rate scheduling,
• Global Average Pooling, and
• Dropout regularization,

contributed significantly to efficient optimization and strong generalization.

---

Classification Report

Class	Precision	Recall	F1-Score	Support
lung_n	0.93	0.93	0.93	164
lung_aca	1.00	1.00	1.00	167
lung_scc	0.93	0.93	0.93	149

Overall Metrics

Metric	Score
Accuracy	0.95
Macro Average	0.95
Weighted Average	0.95

---

Confusion Matrix Analysis

Confusion Matrix

Figure 4: Confusion Matrix.

`

True Class	Correct Predictions	Total Samples	Class Accuracy
lung_n	153	164	93.29%
lung_aca	167	167	100.00%
lung_scc	138	149	92.62%

Key Observations

Lung Adenocarcinoma (lung_aca)

The model achieved perfect classification performance for adenocarcinoma samples:

• 167 / 167 correctly classified
• No false positives
• No false negatives
• Precision = 1.00
• Recall = 1.00
• F1-score = 1.00

This demonstrates exceptionally strong feature extraction capabilities and complete separation from the remaining classes.

Lung Squamous Cell Carcinoma (lung_scc)

The model correctly classified:

138 / 149 samples

Misclassifications:

• 11 samples predicted as normal tissue.
• No confusion with adenocarcinoma samples.

This indicates strong subtype-specific feature learning and robust cancer-class separation.

Normal Lung Tissue (lung_n)

The model correctly classified:

153 / 164 samples

Misclassifications:

• 11 samples predicted as squamous cell carcinoma.
• No samples predicted as adenocarcinoma.

The absence of confusion with adenocarcinoma highlights the effectiveness of the learned feature representations.

Overall Interpretation

The confusion matrix reveals several important characteristics:

• Zero confusion involving adenocarcinoma samples.
• Misclassifications occur exclusively between normal tissue and squamous cell carcinoma.
• Both cancer classes remain distinctly separated.

Achieving perfect adenocarcinoma detection alongside more than 92% class accuracy on the remaining categories demonstrates the effectiveness of residual learning for histopathological image classification.

---

Dataset and Other Files

Dataset: Google Drive Dataset

Model Link Model

Colab File: Colab File

Future Work

Several directions can further enhance this work:

Data Augmentation

Future experiments will explore:

• Random rotations
• Horizontal and vertical flips
• Color jittering
• Elastic transformations
• Stain normalization methods

Increasing sample diversity may further improve robustness and reduce confusion between normal tissue and squamous cell carcinoma.

Advanced Training Strategies

Potential improvements include:

• MixUp and CutMix augmentation.
• Cross-validation experiments.
• Class-balanced sampling.
• Label smoothing techniques.
• Attention modules such as CBAM and SE blocks.
• Transfer learning from large-scale medical imaging models.

Model Interpretability

Future work may also incorporate:

• Grad-CAM visualizations.
• Feature activation analysis.
• Explainable AI (XAI) methods for clinical interpretability.

Deployment

Planned deployment directions include:

• Publishing pretrained weights on Hugging Face.
• Building an interactive inference interface.
• Packaging lightweight deployment pipelines for research and educational use.

---

Tech Stack

• Python
• PyTorch
• Torchvision
• NumPy
• Pandas
• Matplotlib
• Scikit-learn

---

Conclusion

This project demonstrates that a lightweight, custom-designed Residual CNN can effectively classify lung histopathological images with strong generalization and excellent class-wise performance.

By incorporating residual learning, batch normalization, adaptive optimization, global average pooling, and regularization techniques, the model achieved:

• 96.46% peak validation accuracy
• 0.1268 best validation loss
• 100% adenocarcinoma detection (167/167 samples)
• 95% overall classification accuracy
• 0.95 macro and weighted F1-scores

These results highlight the potential of residual architectures for medical image analysis and establish a strong foundation for future work involving larger datasets, advanced augmentation strategies, explainable AI techniques, and deployment-oriented applications.