Spatiotemporal Excitation Module-based CNN for Diagnosis of Hepatic Malignancy in Four-phase CT Images.

Liver cancer is a part of the common causes of cancer death worldwide, and the accurate diagnosis of hepatic malignancy is important for effective next treatment. In this paper, we propose a convolutional neural network (CNN) based on a spatiotemporal excitation (STE) module for identification of hepatic malignancy in four-phase computed tomography (CT) images. To enhance the display detail of lesion, we expand single-channel CT images into three channels by using the channel expansion method. Our proposed STE module consists of a spatial excitation (SE) module and a temporal interaction (TI) module. The SE module calculates the temporal differences of CT slices at the feature level, which is used to excite shape-sensitive channels of the lesion features. The TI module shifts a portion of the channels in the temporal dimension to exchange information among the current CT slice and adjacent CT slices. Four-phase CT images of 398 patients diagnosed with hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICC) are used for experiments and five cross-validations are performed. Our model achieved average accuracy of 85.00% and average AUC of 88.91% for classifying HCC and ICC.Clinical Relevance- The proposed deep learning-based model can perform HCC and ICC recognition tasks based on four-phase CT images, assisting doctors to obtain better diagnostic performance.

Automatic Lung Cancer Subtypes Classification on CT Images with Self-generated Multi-modality Hybrid Features.

Lung cancer is a malignant tumor with rapid progression and high fatality rate. According to histological morphology and cell behaviours of cancerous tissues, lung cancer can be classified into a variety of subtypes. Since different cancer subtype corresponds to distinct therapies, the early and accurate diagnosis is critical for following treatments and prognostic managements. In clinical practice, the pathological examination is regarded as the gold standard for cancer subtypes diagnosis, while the disadvantage of invasiveness limits its extensive use, leading the non-invasive and fast-imaging computed tomography (CT) test a more commonly used modality in early cancer diagnosis. However, the diagnostic results of CT test are less accurate due to the relatively low image resolution and the atypical manifestations of cancer subtypes. In this work, we propose a novel automatic classification model to offer the assistance in accurately diagnosing the lung cancer subtypes on CT images. Inspired by the findings of cross-modality associations between CT images and their corresponding pathological images, our proposed model is developed to incorporate general histopathological information into CT imagery-based lung cancer subtypes diagnostic by omitting the invasive tissue sample collection or biopsy, and thereby augmenting the diagnostic accuracy. Experimental results on both internal evaluation datasets and external evaluation datasets demonstrate that our proposed model outputs more accurate lung cancer subtypes diagnostic predictions compared to existing CT-based state-of-the-art (SOTA) classification models, by achieving significant improvements in both accuracy (ACC) and area under the receiver operating characteristic curve (AUC).Clinical Relevance- This work provides a method for automatically classifying the lung cancer subtypes on CT images.