BioMind影像辅助诊断软件在颅内肿瘤磁共振诊断中的应用

Application of imaging-assisted diagnosis system (BioMind) in the magnetic resonance diagnosis of intracranial tumors

目的:研究BioMind影像辅助诊断软件在颅内肿瘤磁共振诊断中的应用价值。方法:回顾性分析2018年~2021年455例经手术病理确诊颅内肿瘤病例的磁共振影像和病理资料,将符合标准的病例磁共振影像导入到BioMind人工智能影像辅助诊断软件(3.5版本)进行分析,比较此深度学习系统(deep learning system, DLS)和放射科医生诊断的一致性和对总体肿瘤以及不同类型肿瘤的第一诊断和综合诊断的准确率。结果:共283病例资料符合研究标准,包含了14种颅内肿瘤类型。不论第一诊断还是综合诊断,DLS和放射科医生均存在诊断一致性(Kappa值分别为0.331和0.263,P=0.000)。对总体肿瘤,DLS和放射科医生第一诊断准确率差异无统计学意义(83.4%vs 86.9%,P=0.194),综合诊断准确率差异无统计学意义(91.9%vs 90.8%,P=0.728);对脑膜瘤(160例),DLS第一诊断准确率低于放射科医生(91.2%vs97.5%,P=0.021),综合诊断准确率差异无统计学意义(96.2%vs 98.8%,P=0.728);对垂体腺瘤(47例),DLS和放射科医生第一诊断准确率差异无统计学意义(95.7%vs100%,P=0.500),综合诊断准确率差异无统计学意义(95.7%vs100%,P=0.500);对胶质母细胞瘤(30例),DLS和放射科医生第一诊断准确率无显著差异(76.7%vs63.3%,P=0.372),综合诊断准确率DLS高于放射科医生(100%vs76.7%,P=0.016)。对星形细胞瘤(22例),DLS第一诊断准确率高于放射科医生(45.4%vs40.9%,P=0.027),综合诊断准确率高于放射科医生(63.6%vs45.4%,P=0.024)。结论:基于磁共振影像数据,BioMind人工智能影像辅助诊断软件在颅内肿瘤分类方面具有较高的准确性,可以辅助医生提高诊断效率,具有较好的推广应用前景。

Objective:To study the application value of BioMind artificial intelligence imaging-assisted diagnosis system in the magnetic resonance diagnosis of intracranial tumors. Methods: Magnetic Resonance(MR) images and pathological data of 367 cases of intracranial tumors confirmed by surgical pathology in our hospital from 2018-2021 were retrospectively analyzed. The MR images of the cases included in the study were imported into the BioMind(Version3.5) artificial intelligence imaging-assisted diagnosis system for analysis, and the accuracy of the first diagnosis and comprehensive diagnosis of the deep learning system(DLS) and radiologists were counted and compared on all tumors and different types of tumors. Results: A total of 283 cases with 14 types of intracranial tumors were included in the study. For both the first and comprehensive diagnosis, agreement was observed between DLS and radiologists(kappa coefficient was 0.331 and 0.263, P=0.000). For all tumors, there was no significant difference in the first diagnostic accuracy between the DLS and radiologists(83.4% vs. 86.9%, P=0.194) and no significant difference in the comprehensive diagnostic accuracy(91.9% vs. 90.8%, P=0.728);For meningiomas(160 cases), the first diagnostic accuracy of the DLS was lower than radiologists(91.2% vs. 97.5%,P=0.021) and the comprehensive accuracy was not significantly different(96.2% vs. 98.8%, P=0.728). For pituitary adenoma(47 cases), the first and comprehensive diagnostic accuracy had no significant difference between the DLS and radiologists(95.7% vs. 100%, P=0.500;95.7% vs. 100%, P=0.500). For glioblastoma(30 cases), there was no significant difference in the first diagnostic accuracy between the DLS and radiologists(76.7% vs. 63.3%, P=0.372), the comprehensive accuracy of the DLS was higher than radiologists(100% vs. 76.7%, P=0.016). For astrocytoma(22 cases), both the first and comprehensive diagnostic accuracy of the DLS was higher than radiologists(45.4% vs. 40.9%, P=0.027;63.6% vs. 45.4%, P=0.024). Conclusion:Based on MR image data, the deep learning system has high accuracy in intracranial tumor classification, which can assist radiologists to improve the diagnostic efficiency and has a superior prospect for promotion and application.