Identification lymph node metastasis in esophageal squamous cell carcinoma using whole slide images and a hybrid network of multiple instance and transfer learning. (April 2023)

• Record Type: Journal Article
• Authors: Kang, Huan; Yang, Meilin; Zhang, Fan; Xu, Huiya; Ren, Shenghan; Li, Jun; Chen, Duofang; Wang, Fen; Li, Dan; Chen, Xueli
• Is Part Of: Biomedical signal processing and control. Volume 82(2023)
• Publication Date: 2023-04
• Subjects: Lymph node metastasis -- Esophageal squamous cell carcinoma -- Weakly-supervised learning -- Whole slide image -- Classification; EC esophageal cancer -- ESCC esophageal squamous cell carcinoma -- LNM lymph node metastasis -- ESCC LNM lymph node metastasis of esophageal squamous cell carcinoma -- Breast LNM lymph node metastasis of breast cancer -- WSI whole slide image -- H&E hematoxylin and eosin -- TCGA The Cancer Genome Atlas -- MITL the target model proposed by ourself which is pre-trained the aggregation network on the dataset of Breast LNM and then fine-tuned the parameters on the dataset of ESCC LNM -- B-MIL the pre-training model based on MIL which is trained by the dataset of Breast LNM -- E-MIL the direct prediction model based on MIL which is trained by the dataset of ESCC LNM -- SL-MITL the feature extractor in MITL is replaced by an in-domain ResNet50 -- SL-E-MIL the feature extractor in E-MIL is replaced by an in-domain ResNet50 -- ACC accuracy -- AUC the score of area under the curve -- MCC Matthews correlation coefficient -- Kappa Kappa coefficient; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2023.104577
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: MITL was designed to accurately identify ESCC LNM using slide-level labels. MITL offers astounding performance under the low demand of ESCC LNM dataset. The AUC reached 0.988, is 0.028 higher than the published using pixel-level labels. Pre-training the aggregation network has more advantages than the feature extractor. Abstract: Difficulties associated with identifying lymph nodes metastasis in esophageal squamous cell carcinoma (ESCC LNM) can make it challenging to determine the clinical stage and devize precise treatment strategies for patients with esophageal cancer (EC). The lack of a large public dataset and expensive expert annotation are the factors responsible for the slow development of clinical computer-aided diagnosis for ESCC LNM. In this study, we collected 863 whole slide images from 198 patients at two hospitals, and developed a weakly supervised workflow based on a hybrid network of multiple instance and transfer learning (MITL) for automating identification of ESCC LNM. The results showed that MITL achieved a significant performance advantage over its competitors. The accuracy (ACC), F1-score, Matthews correlation coefficient, and areas under the curve were 0.976, 0.944, 0.929, and 0.991 for the internal testing data, and 0.969, 0.925, 0.905, and 0.988 for the external testing data, respectively. Compared to pre-train feature extractor, the improvement in ACC of pre-training aggregation network on the internal testing data was approximately twice. Furthermore, ACCs were determined using MITL from micrometastasis and macrometastasis (0.824 and 0.95, respectively). Visualization showed that the key features of ESCC LNM can be extracted by MITL for accurate detection and classification. In summary, our findings illustrated that MITL can achieve the high-efficiency identification and prediction of ESCC LNM with less investment of resources, and provide a new research strategy for the diagnosis of it.