[Paper Review] Comparison of different CNNs for breast tumor classification from ultrasound images

Jorge F. Lazo, Sara Moccia+2 more|arXiv (Cornell University)|Dec 28, 2020

AI in cancer detection9 citations

TL;DR

This study evaluates fine-tuning and feature extraction using pre-trained VGG-16 and Inception V3 models for classifying breast tumors in ultrasound images. Fine-tuning VGG-16 achieved the highest performance with 91.9% accuracy and 0.934 AUC, outperforming other configurations and highlighting fine-tuning as the superior strategy for this task with limited data.

ABSTRACT

Breast cancer is one of the deadliest cancer worldwide. Timely detection could reduce mortality rates. In the clinical routine, classifying benign and malignant tumors from ultrasound (US) imaging is a crucial but challenging task. An automated method, which can deal with the variability of data is therefore needed. In this paper, we compared different Convolutional Neural Networks (CNNs) and transfer learning methods for the task of automated breast tumor classification. The architectures investigated in this study were VGG-16 and Inception V3. Two different training strategies were investigated: the first one was using pretrained models as feature extractors and the second one was to fine-tune the pre-trained models. A total of 947 images were used, 587 corresponded to US images of benign tumors and 360 with malignant tumors. 678 images were used for the training and validation process, while 269 images were used for testing the models. Accuracy and Area Under the receiver operating characteristic Curve (AUC) were used as performance metrics. The best performance was obtained by fine tuning VGG-16, with an accuracy of 0.919 and an AUC of 0.934. The obtained results open the opportunity to further investigation with a view of improving cancer detection.

Motivation & Objective

• To evaluate the effectiveness of transfer learning strategies—feature extraction and fine-tuning—on pre-trained CNNs for breast tumor classification in ultrasound images.
• To compare the performance of two widely used architectures, VGG-16 and Inception V3, under different training strategies on a limited dataset.
• To identify the optimal combination of model architecture and training method for maximizing classification accuracy and AUC in automated breast tumor diagnosis.
• To investigate overfitting tendencies in models trained on small medical imaging datasets and assess the need for regularization techniques.

Proposed method

• Pre-trained VGG-16 and Inception V3 models were used as the base networks for transfer learning.
• Two training strategies were applied: (1) freezing all base layers and using the network solely as a feature extractor, and (2) fine-tuning the last few layers of the pre-trained models.
• A new fully connected (FC) layer with 1024 (VGG-16) or 512 (Inception V3) neurons was added on top of the feature extractor for classification.
• Training was performed with a base learning rate of 0.001, a batch size of 50, and the Adam optimizer on an NVIDIA GTX 1660 GPU using Keras with TensorFlow backend.
• Data augmentation was not explicitly applied, but normalization techniques like dropout and L2 regularization were considered to mitigate overfitting.
• Performance was evaluated using accuracy and area under the ROC curve (AUC), with 678 images used for training/validation and 269 for testing.

Experimental results

Research questions

• RQ1 Which transfer learning strategy—feature extraction or fine-tuning—yields better performance for breast tumor classification in ultrasound images?
• RQ2 Does the choice of pre-trained CNN architecture (VGG-16 vs. Inception V3) significantly affect classification performance on a limited dataset?
• RQ3 How do the performance metrics (accuracy and AUC) vary between models during training and testing, and what does this suggest about overfitting?
• RQ4 Can fine-tuned VGG-16 generalize effectively on a small, variable dataset of breast ultrasound images?

Key findings

• Fine-tuning VGG-16 achieved the highest test accuracy of 0.919 and AUC of 0.934, outperforming all other configurations.
• The feature extraction strategy yielded significantly lower performance, with VGG-16 achieving only 0.862 accuracy and 0.791 AUC.
• Inception V3 showed poor generalization, achieving 0.756 accuracy and 0.783 AUC when fine-tuned, despite reaching over 0.93 accuracy during training, indicating overfitting.
• The performance gap between fine-tuning and feature extraction was larger for AUC (0.16) than for accuracy (0.046), suggesting better discrimination capability with fine-tuning.
• Overfitting was observed in Inception V3 during training, with high training accuracy (over 0.93) but lower test performance, highlighting the need for regularization.
• Visual analysis of predictions revealed that image variability significantly affects model generalization, especially for benign tumors misclassified as malignant.