Fast Convolutional Neural Network Training Using Selective Data Sampling: Application to Hemorrhage Detection in...

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) are advanced deep learning models widely used in computer vision and medical image analysis.
• Training CNNs, especially for medical imaging, is often lengthy and inefficient due to a large number of easily classifiable samples.

Purpose of the Study:

• To propose and evaluate a novel method for improving and accelerating CNN training in medical image analysis.
• To enhance the efficiency of CNN learning by dynamically selecting informative training samples.

Main Methods:

• A selective sampling (SeS) method was developed to dynamically select misclassified negative samples during CNN training.
• Training samples are assigned weights, prioritizing informative samples for subsequent training iterations.
• The proposed method was compared against a standard training approach (NSeS) using CNNs for hemorrhage detection in fundus images.

Main Results:

• The selective sampling method (SeS) reduced CNN training time from 170 to 60 epochs.
• Performance, measured by area under the receiver operating characteristics curve, reached levels comparable to human experts (0.894 and 0.972 on two datasets).
• The SeS CNN demonstrated statistically superior performance over the NSeS CNN on an independent test set.

Conclusions:

• The proposed selective sampling method significantly speeds up CNN training for medical image analysis tasks.
• This approach enhances CNN performance and efficiency, achieving expert-level accuracy in hemorrhage detection.
• Selective sampling is a promising technique for optimizing deep learning in medical imaging applications.