Anagnostopoulos Ioannis, Maglogiannis Ilias
Department of Information and Communication Systems Engineering, University of the Aegean, Karlovassi, 83200, Samos, Greece.
Med Biol Eng Comput. 2006 Sep;44(9):773-84. doi: 10.1007/s11517-006-0079-4. Epub 2006 Aug 3.
This paper deals with breast cancer diagnostic and prognostic estimations employing neural networks over the Wisconsin Breast Cancer datasets, which consist of measurements taken from breast cancer microscopic instances. A probabilistic approach is dedicated to solve the diagnosis problem, detecting malignancy among instances derived from the Fine Needle Aspirate test, while regression algorithms estimate the time interval that possibly correspond to the right end-point of the patients' disease-free survival time or the time where the tumour recurs (time-to-recur). For the diagnosis problem, the accuracy of the neural network in terms of sensitivity and specificity was measured at 98.6 and 97.5% respectively, using the leave-one-out test method. As far as the prognosis problem is concerned, the accuracy of the neural network was measured through a stratified tenfold cross-validation approach. Sensitivity ranged between 80.5 and 91.8%, while specificity ranged between 91.9 and 97.9%, depending on the tested fold and the partition of the predicted period. The prognostic recurrence predictions were then further evaluated using survival analysis and compared with other techniques found in literature.
本文利用威斯康星乳腺癌数据集，运用神经网络进行乳腺癌的诊断和预后评估，该数据集包含从乳腺癌微观实例中获取的测量数据。采用概率方法解决诊断问题，在细针穿刺抽吸测试所得实例中检测恶性肿瘤，而回归算法则估计可能对应患者无病生存期右端点或肿瘤复发时间（复发时间）的时间间隔。对于诊断问题，使用留一法测试方法，神经网络在敏感性和特异性方面的准确率分别为98.6%和97.5%。就预后问题而言，通过分层十折交叉验证方法测量神经网络的准确率。敏感性在80.5%至91.8%之间，特异性在91.9%至97.9%之间，具体取决于测试折数和预测期的划分。然后使用生存分析进一步评估预后复发预测，并与文献中发现的其他技术进行比较。
