Infrared imaging has been used for detecting early-stage breast abnormalities. However, manual interpretation of these images is subjective and error prone. Automated analysis of thermal images using artificial intelligence increases the accuracy of detecting breast cancer and enables use in breast cancer screening programmes. A crucial step in automated analysis is tumour segmentation, which is the focus of this paper. Prior algorithms used for tumour segmentation involve clustering, thresholding and active contour techniques that rely on low-level image features generated from manual feature engineering. The recent advances in deep learning, especially convolutional neural networks (CNNs), are slowly making them the de-facto method for automated image analysis. In this paper, we explore various CNN architectures for semantic segmentation starting from naive patch-based classifiers to more sophisticated ones such as encoder-decoder architecture for detecting the hotspots in the thermal image. We also show that encoder-decoder architectures perform better when compared to patch-based classifiers in terms of accuracy, Dice index, Jaccard index and inference time even with a smaller dataset.

红外成像已用于检测早期乳房异常。但是，手动解释这些图像是主观的并且容易出错。使用人工智能对热图像进行自动分析可以提高检测乳腺癌的准确性，并可以用于乳腺癌筛查程序。自动分析的关键步骤是肿瘤分割，这是本文的重点。用于肿瘤分割的现有算法涉及聚类，阈值化和主动轮廓技术，这些技术依赖于从手动特征工程生成的低级图像特征。深度学习的最新进展，尤其是卷积神经网络（CNN），正在使它们逐渐成为自动化图像分析的事实上的方法。在本文中，我们探索了用于语义分割的各种CNN架构，从基于朴素补丁的分类器到更复杂的分类器，例如用于检测热图像中热点的编码器-解码器架构。我们还表明，与基于补丁的分类器相比，即使在数据集较小的情况下，编码器-解码器体系结构在准确性，Dice索引，Jaccard索引和推理时间方面也表现更好。