[Paper Review] Can Pretrained Neural Networks Detect Anatomy?

Vlado Menkovski, Zharko Aleksovski+2 more|arXiv (Cornell University)|Dec 18, 2015

AI in cancer detection7 references3 citations

TL;DR

This paper investigates transfer learning and data augmentation for anatomy detection in X-ray images using pretrained and custom convolutional neural networks. It demonstrates that fine-tuning a pretrained OverFeat network with SVM achieves 92.42% accuracy, while a custom CNN with heavy data augmentation and regularization reaches 95.14%, showing that robust feature learning is possible even with limited medical data when proper regularization and augmentation are applied.

ABSTRACT

Convolutional neural networks demonstrated outstanding empirical results in computer vision and speech recognition tasks where labeled training data is abundant. In medical imaging, there is a huge variety of possible imaging modalities and contrasts, where annotated data is usually very scarce. We present two approaches to deal with this challenge. A network pretrained in a different domain with abundant data is used as a feature extractor, while a subsequent classifier is trained on a small target dataset; and a deep architecture trained with heavy augmentation and equipped with sophisticated regularization methods. We test the approaches on a corpus of X-ray images to design an anatomy detection system.

Motivation & Objective

• To address the challenge of limited annotated medical imaging data by exploring transfer learning and data augmentation strategies.
• To evaluate whether pretrained neural networks trained on natural images can generalize to medical anatomy detection.
• To investigate if deep networks can be effectively trained on small medical datasets using strong regularization and augmentation.
• To compare the performance of off-the-shelf feature extractors (e.g., OverFeat) with custom architectures trained end-to-end on medical data.

Proposed method

• Fine-tuning a pretrained OverFeat network on ImageNet to extract 4096-dimensional features from X-ray images.
• Training a linear multi-class SVM on the extracted features using 5-fold cross-validation and exhaustive hyperparameter search for regularization (C).
• Designing a custom deep CNN inspired by Razavian et al., using batch normalization, dropout, and leaky ReLU activations.
• Applying extensive data augmentation including resizing, cropping, rotation, translation, shearing, stretching, and flipping to increase effective training data.
• Training the custom CNN end-to-end on a small X-ray dataset of 14,676 images with 24 anatomical classes.
• Evaluating both approaches using a 90/10 train-test split on the ImageClef 2009 medical image dataset.

Experimental results

Research questions

• RQ1 Can a neural network pretrained on natural images serve as an effective feature extractor for anatomy detection in X-ray images?
• RQ2 How does performance compare between transfer learning using a pretrained network and training a custom deep network from scratch with data augmentation?
• RQ3 To what extent can data augmentation and regularization compensate for limited annotated medical training data?
• RQ4 Is fine-tuning only the last layer of a pretrained network sufficient for effective transfer learning in medical imaging?

Key findings

• The OverFeat + SVM approach achieved a test accuracy of 92.42% on the X-ray anatomy detection task.
• The custom CNN trained with heavy data augmentation and regularization reached a higher accuracy of 95.14%, outperforming the transfer learning baseline.
• Pretrained networks like OverFeat serve as strong generic image descriptors even when applied outside their original training domain.
• Sufficiently deep architectures can be successfully trained on small medical datasets when combined with aggressive data augmentation and regularization techniques.
• Fine-tuning only the last layer of a pretrained network is a viable and effective transfer learning strategy in data-scarce medical imaging scenarios.