Fine-Tuning Convolutional Neural Networks for Biomedical Image Analysis: Actively and Incrementally

Abstract

Intense interest in applying convolutional neural networks (CNNs) in biomedical image analysis is wide spread, but its success is impeded by the lack of large annotated datasets in biomedical imaging. Annotating biomedical images is not only tedious and time consuming, but also demanding of costly, specialty-oriented knowledge and skills, which are not easily accessible. To dramatically reduce annotation cost, this paper presents a novel method called AIFT (active, incremental fine-tuning) to naturally integrate active learning and transfer learning into a single framework. AIFT starts directly with a pre-trained CNN to seek “worthy” samples from the unannotated for annotation, and the (fine-tuned) CNN is further fine-tuned continuously by incorporating newly annotated samples in each iteration to enhance the CNN’s performance incrementally. We have evaluated our method in three different biomedical imaging applications, demonstrating that the cost of annotation can be cut by at least half. This performance is attributed to the several advantages derived from the advanced active and incremental capability of our AIFT method.

Learning Loss for Active Learning

“…As the two approaches based on uncertainty and distribution are differently motivated, they are complementary to each other. Thus, a variety of hybrid strategies have been proposed [29,59,41,56] for their specific tasks.…”

Exploiting the potential of unlabeled endoscopic video data with self-supervised learning

“…While approaches to semi-supervised learning, which typically handle unlabeled and labeled data from the same data distribution simultaneously, are increasingly common in the field of Medical Image Computing (MIC) [2], we are, to our knowledge, the first to investigate the concept of self-supervised learning to reduce manual labeling effort in medical image segmentation. In contrast to state-of-the-art pre-training methods [24,27,28,32], we initialized our model on the target domain using only unlabeled data rather than on a different domain with labeled data. This is achieved with an auxiliary task that can be assumed to learn a representation of the target domain that is well-suited for the target task.…”

Cost-Effective Training of Deep CNNs with Active Model Adaptation

“…Note that there is no general approach applicable to our setting, we compare the proposed approach with the Random method, which randomly select instances to query their labels. The method AIFT proposed in [44] was originally designed for binary classification of medical images, and thus is not compared in the multi-class cases. Instead, the performance of the fully re-trained model is provided for reference.…”