Learning to Segment Skin Lesions from Noisy Annotations

Abstract

Deep convolutional neural networks have driven substantial advancements in the automatic understanding of images. Requiring a large collection of images and their associated annotations is one of the main bottlenecks limiting the adoption of deep networks. In the task of medical image segmentation, requiring pixel-level semantic annotations performed by human experts exacerbate this difficulty. This paper proposes a new framework to train a fully convolutional segmentation network from a large set of cheap unreliable annotations and a small set of expert-level clean annotations. We propose a spatially adaptive reweighting approach to treat clean and noisy pixel-level annotations commensurately in the loss function. We deploy a meta-learning approach to assign higher importance to pixels whose loss gradient direction is closer to those of clean data. Our experiments on training the network using segmentation ground truth corrupted with different levels of annotation noise show how spatial reweighting improves the robustness of deep networks to noisy annotations.

Embracing imperfect datasets: A review of deep learning solutions for medical image segmentation

“…It is therefore important to utilize strategies that mitigate the adverse effects of label noise during training. Mirikharaji et al (2019) propose a learning algorithm resilient to the label noise in the segmentation masks. The suggested method consists of a weighted cross entropy loss function where the contribution of each pixel to the total loss is controlled by model's perception of the annotation quality for the pixels.…”

“…Training a classification model with aggregation layers and using activation maps as class-specific segmentation annotation variability on the trained model, but it also enables training with only rough annotations, which can be obtained in a cost-effective manner with significantly shorter annotation time than that of accurate annotations. For instance, the work by Mirikharaji et al (2019) shows that, with a noise-resilient approach, a skin segmentation model trained with 3-vertex contours can achieve similar performance to a model trained using accurate segmentation masks. Handling annotation noise in medical segmentation datasets is still a fairly new topic and deserves further investigation.…”

A Noise-Robust Framework for Automatic Segmentation of COVID-19 Pneumonia Lesions From CT Images

“…To deal with noisy labels for segmentation of medical images, a CNN was designed in [15] to distinguish noisy labels from clean ones. In [19], metalearning was used to assign lower weights to pixels whose loss gradient direction is further from those of clean data. Both methods require a set of clean labels for training.…”

Meta-Learning in Neural Networks: A Survey

“…In [275] MAML is adapted to weaklysupervised breast cancer detection tasks, and the order of tasks are selected according to a curriculum. MAML is also combined with denoising autoencoders to do medical visual question answering [276], while learning to weigh support samples [218] is adapted to pixel wise weighting for skin lesion segmentation tasks that have noisy labels [277].…”