Barua Barun, Chyrmang Genevieve, Bora Kangkana, Saikia Manob Jyoti
Department of Computer Science and Information Technology, Cotton University, Guwahati, Assam, India.
Electrical and Computer Engineering Department, University of Memphis, Memphis, TN, United States of America.
PeerJ Comput Sci. 2024 Dec 23;10:e2633. doi: 10.7717/peerj-cs.2633. eCollection 2024.
Colorectal cancer represents a significant health challenge as one of the deadliest forms of malignancy. Manual examination methods are subjective, leading to inconsistent interpretations among different examiners and compromising reliability. Additionally, process is time-consuming and labor-intensive, necessitating the development of computer-aided diagnostic systems. This study investigates the segmentation of colorectal cancer regions of normal tissue, polyps, high-grade intraepithelial neoplasia, low-grade intraepithelial neoplasia, adenocarcinoma, and serrated Adenoma, using proposed segmentation models: VGG16-UNet, ResNet50-UNet, MobileNet-UNet, and MobileViT-UNet. This is the first study to integrate MobileViT as a UNet encoder. Each model was trained with two distinct loss functions, binary cross-entropy and dice loss, and evaluated using metrics including Dice ratio, Jaccard index, precision, and recall. The MobileViT-UNet+Dice loss emerged as the leading model in colorectal histopathology segmentation, consistently achieving high scores across all evaluation metrics. Specifically, it achieved a Dice ratio of 0.944 ± 0.030 and a Jaccard index of 0.897 ± 0.049, with precision at 0.955 ± 0.046 and Recall at 0.939 ± 0.038 across all classes. To further obtain the best performing model, we employed multi-criteria decision analysis (MCDA) using the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS). This analysis revealed that the MobileViT-UNet+Dice model achieved the highest TOPSIS scores of 1, thereby attaining the highest ranking among all models. Our comparative analysis includes benchmarking with existing works, the results highlight that our best-performing model (MobileViT-UNet+Dice) significantly outperforms existing models, showcasing its potential to enhance the accuracy and efficiency of colorectal cancer segmentation.
结直肠癌作为最致命的恶性肿瘤形式之一，是一项重大的健康挑战。手动检查方法具有主观性，导致不同检查人员的解读不一致，降低了可靠性。此外，该过程耗时且 labor-intensive，因此需要开发计算机辅助诊断系统。本研究使用提出的分割模型：VGG16-UNet、ResNet50-UNet、MobileNet-UNet 和 MobileViT-UNet，对正常组织、息肉、高级别上皮内瘤变、低级别上皮内瘤变、腺癌和锯齿状腺瘤的结直肠癌区域进行分割。这是第一项将 MobileViT 集成作为 UNet 编码器的研究。每个模型都使用两种不同的损失函数（二元交叉熵和骰子损失）进行训练，并使用包括骰子比率、杰卡德指数、精度和召回率在内的指标进行评估。MobileViT-UNet+骰子损失在结直肠癌组织病理学分割中成为领先模型，在所有评估指标上始终获得高分。具体而言，它在所有类别中的骰子比率为 0.944 ± 0.030，杰卡德指数为 0.897 ± 0.049，精度为 0.955 ± 0.046，召回率为 0.939 ± 0.038。为了进一步获得性能最佳的模型，我们使用与理想解相似性排序法（TOPSIS）进行多准则决策分析（MCDA）。该分析表明，MobileViT-UNet+骰子模型的 TOPSIS 得分最高，为 1，从而在所有模型中排名最高。我们的比较分析包括与现有工作进行基准测试，结果表明我们性能最佳的模型（MobileViT-UNet+骰子）明显优于现有模型，展示了其提高结直肠癌分割准确性和效率的潜力。
