基于ResNet50和Grad-CAM的卵巢病变分类模型：深度学习在医学影像诊断中的应用

Ovarian Lesion Classification Model Based on ResNet50 and Grad-CAM: Application of Deep Learning in Medical Imaging Diagnosis

摘要: 卵巢病变的早期诊断对治疗和预后至关重要，但传统诊断方法存在解读差异。本研究提出了一种基于ResNet50深度学习网络的卵巢病变分类模型，结合Grad-CAM技术生成热力图，以增强模型的可解释性。通过对930张超声图像进行五折交叉验证，模型在区分正常、恶性、畸胎瘤、良性肿瘤交界性和巧克力囊肿等病理类型方面表现出较高的准确性(平均AUC为88.16)。热力图可视化显示，模型能够有效识别病变区域，特别是在恶性肿瘤和畸胎瘤的分类中表现出色。研究表明，该模型在卵巢病变的早期筛查和辅助诊断中具有重要应用价值，未来可进一步优化以应对复杂病例。

Abstract: Early diagnosis of ovarian lesions is crucial for treatment and prognosis, but traditional diagnostic methods are subject to interpretation variability. This study proposes an ovarian lesion classification model based on the ResNet50 deep learning network, combined with Grad-CAM technology to generate heatmaps, enhancing the interpretability of the model. Through five-fold cross-validation on 930 ultrasound images, the model demonstrated high accuracy in distinguishing pathological types such as normal, malignant, teratoma, borderline benign tumors, and chocolate cysts (average AUC of 88.16). Heatmap visualization revealed that the model effectively identified lesion areas, particularly excelling in the classification of malignant tumors and teratomas. The study indicates that this model has significant application value in the early screening and auxiliary diagnosis of ovarian lesions, and future work can further optimize the model to address complex cases.