U-CatcHCC: An Accurate HCC Detector in Hepatic DCE-MRI Sequences Based on an U-Net Framework

U-Net and Its Variants for Medical Image Segmentation: A Review of Theory and Applications

“…Cancer is a leading cause of death worldwide, and MR is one of the strongest methods for the proper prognosis of different types of cancers. In addition to brain cancer, we have found applications on prostate cancer [58], [64], [152]- [156], liver cancer [21], [157], [158], nasopharyngeal cancer [25], [98], [159], and breast cancer [99], [160]. Other implementations include segmentation of the femur [12]- [14], spinal cord [161], [162], blood vessels [100], vertebral column [17], human placenta [163], and the uterus [164].…”

A Novel Deep Learning Approach for Liver MRI Classification and HCC Detection

“…The total accuracy level reached 90% in both phases. Regarding our previous studies [6] and [8], F1 score level has been increased and reached respectively 74 % and 73 % for phase 2 and phase 3 considering [6].This rate is the most suitable and adapted measure to evaluate our model because the used classes have not the same size. In fact, F1 score has been increased from 66.5% in [8] to 74% in our work.…”

“…The proposed model provides better classification results. Results obtained by our model were illustrated respectively in tables 3 and 4 for both phases 2 (With Contrast) and 3 (After contrast).To evaluate our model, we have used the same datasets employed in works [6,8].We have applied the SVM method in [6] and U-Net in [8] while we have applied CNN in this work. The performance of the proposed CNN model is determined by its ability to detect cancerous and normal patches from MRI images.…”

“…This classification approach is realized through different steps: a) crop and resize a set of images around each tumor, b) split randomly lesions into train and test sets, c) create 100 copies of each image presenting a lesion through various image processing techniques, d) using the training set to associate image patterns with the adequate lesion class, e) validation of the model by using the test set and f) evaluating the model performance. In addition, [6]integrated deep learning into MRI image and proposed a framework for HCC lesions detection from DCE-MRI (Dynamic Contrast-Enhanced MRI) sequences. This approach has employed the U-Net architecture in the classification process.…”

“…It takes as input MRI images and gives fragments with equal sizes as output, and 2) creates a parallel CNN algorithm that allows HCC detection and localization in MRI images. The goal here is to increase the Dice (or F-measure, or F1) compared to our previous works [6,8] by trying to keep an interesting accuracy rate. Dice coefficient is considered as an important measure in medical imaging.…”