Multiple instance learning (MIL) models have achieved remarkable success in analyzing whole slide images (WSIs) for disease classification problems. However, with regard to gigapixel WSI classification problems, current MIL models are often incapable of differentiating a WSI with extremely small tumor lesions. This minute tumor-to-normal area ratio in a MIL bag inhibits the attention mechanism from properly weighting the areas corresponding to minor tumor lesions. To overcome this challenge, we propose salient instance inference MIL (SiiMIL), a weakly-supervised MIL model for WSI classification. Our method initially learns representations of normal WSIs, and it then compares the normal WSIs representations with all the input patches to infer the salient instances of the input WSI. Finally, it employs attention-based MIL to perform the slide-level classification based on the selected patches of the WSI. Our experiments imply that SiiMIL can accurately identify tumor instances, which could only take up less than 1% of a WSI, so that the ratio of tumor to normal instances within a bag can increase by two to four times. It is worth mentioning that it performs equally well for large tumor lesions. As a result, SiiMIL achieves a significant improvement in performance over the state-of-the-art MIL methods.
翻译:多重实例学习模型(MIL)在分析用于疾病分类问题的整张幻灯片图像(WSI)方面取得了显著的成功。然而,关于千兆像素 WSI分类问题,目前的MIL模型往往无法区分具有极小肿瘤损伤的WSI。MIL包中的这个微小肿瘤对正常区域比率使关注机制无法适当权衡与轻微肿瘤损伤相应的区域。为了克服这一挑战,我们建议了突出的例子推介MIL(SiiMIL),这是一个微弱监督的用于WSI分类的MIL模型。我们的方法最初学习了正常的 WSI的表示方式,然后将正常的WSI表示方式与所有输入的补补丁进行比较,以推断输入的WSI的显著损伤。最后,它使用基于注意的MIL来根据WSI的选定补丁进行幻灯片等级分类。我们的实验表明,SiiMIL能够准确地辨别肿瘤的发生情况,这只能占WSI的1%以下,因此包内肿瘤与正常情况的比率可以增加2至4倍。SIMI的性能显著改善。