Tumor Classification Using an Ensemble of ShuffleNet_V2, MobileNet_V2, and ResNet18

---

Abstract

Breast cancer remains one of the leading causes of mortality among women worldwide.
This project presents a lightweight ensemble deep learning model that classifies tumors as benign or malignant from mammographic images.

Three efficient convolutional neural networks—ShuffleNet_V2, MobileNet_V2, and ResNet18—are trained independently and combined using a bagging ensemble technique.
The ensemble averages their softmax outputs to form robust predictions.

To improve interpretability, the pipeline also visualizes side-by-side Craniocaudal (CC) and Mediolateral Oblique (MLO) projections during validation.

---

1. Introduction

Breast cancer screening typically uses two standard mammogram views—CC and MLO.
Given the subtle differences in tissue patterns, automated classification can be difficult.
By combining the strengths of multiple CNNs, this project enhances prediction accuracy while maintaining computational efficiency.

---

2. Dataset

The dataset contains mammographic images labeled as benign or malignant.
Each sample includes both CC and MLO images, offering complementary perspectives of breast tissue.

---

🧩 3. Model Architecture Overview

3.1 ShuffleNet_V2

• Key Features:
  • Channel Shuffle for inter-group feature communication.
  • Grouped Convolutions to reduce computational cost.
  • Lightweight design suited for deployment on edge devices.
• Contribution to Ensemble: Extracts high-level features efficiently.

---

3.2 MobileNet_V2

• Key Features:
  • Depthwise Separable Convolutions reduce parameters dramatically.
  • Inverted Residuals with linear bottlenecks for compact yet expressive modeling.
• Contribution to Ensemble: Balances efficiency and accuracy with mid-level feature extraction.

---

3.3 ResNet18

• Key Features:
  • Skip Connections to avoid vanishing gradients.
  • 17 Convolutional Layers + 1 FC layer for deeper representation learning.
  • The 18 in ResNet-18 refers to the no. of layers that have learnable parameters i.e 17 conv layers and 1 FC layer
• Contribution to Ensemble: Captures complex hierarchical features, complementing lighter models.

---

⚙️ 4. Ensemble Model Formation

Each model is trained individually. During inference, the ensemble prediction is computed as follows:

1. Each model outputs a probability distribution over the classes using softmax.

2. The outputs are averaged element-wise:

   $$ P_{\text{final}} = \frac{1}{3} \sum_{i=1}^{3} P_i $$

3. The predicted class is chosen as:

   $$ \hat{y} = \arg\max (P_{\text{final}}) $$

This represents a bagging ensemble strategy, reducing variance and improving generalization.

---

5. Implementation

The project was implemented using Python 3.12.8 and PyTorch 2.6.0 on PyCharm Professional.

Key Steps:

1. Dataset Preparation:
   Each mammogram sample contains both CC and MLO views, preprocessed with resizing, augmentation, and normalization.

2. Training Setup:

   • Epochs: 10
   • Optimizer: Adam (learning rate = 1e-4)
   • Loss: Cross-Entropy
   • Batch Size: 16
   • Device: CUDA or CPU automatically detected

3. Visualization:
   • Displays side-by-side CC and MLO images with predictions and ground truths.
   • Generates a confusion matrix and classification report post-validation.

6. Evaluation Metrics

6.1 Confusion Matrix

Actual / Predicted	Benign	Malignant
Benign	15	5
Malignant	5	25

6.2 Classification Report

Class	Precision	Recall	F1-score	Support
Benign	0.75	0.75	0.75	20
Malignant	0.83	0.83	0.83	30
Accuracy	0.80			50
Macro avg	0.79	0.79	0.79	50
Weighted avg	0.80	0.80	0.80	50

Metric Definitions

• Precision:
  $$ \text{Precision} = \frac{TP}{TP + FP} $$

• Recall:
  $$ \text{Recall} = \frac{TP}{TP + FN} $$

• F1-Score:
  $$ F_1 = 2 \times \frac{\text{Precision} \times \text{Recall}}{\text{Precision} + \text{Recall}} $$

---

7. Conclusion

The ensemble of ShuffleNet_V2, MobileNet_V2, and ResNet18 effectively enhances tumor classification accuracy from mammograms.
This hybrid approach integrates:

• The efficiency of ShuffleNet,
• The balance of MobileNet, and
• The depth of ResNet.

Visual interpretability using CC and MLO projections and robust evaluation metrics confirm the clinical applicability of the approach.
Future work can explore data augmentation and transfer learning with larger datasets for even better generalization.

---