School of Computer Science and Engineering, Nanjing University of Science and Technology, China.
School of Information Science and Engineering, University of Jinan, China.
Med Image Anal. 2021 Feb;68:101893. doi: 10.1016/j.media.2020.101893. Epub 2020 Nov 12.
The automated prediction of geographic atrophy (GA) lesion growth can help ophthalmologists understand how the GA progresses, and assess the efficiency of current treatment and the prognosis of the disease. We developed an integrated time adaptive prediction model for identifying the location of future GA growth. The proposed model was comprised of bi-directional long short-term memory (BiLSTM) network-based prediction module and convolutional neural network (CNN)-based refinement module. Considering the discontinuity of time intervals among sequential follow-up visits, we integrated time factors into BiLSTM-based prediction module to control the time attribute expediently. Then, the results from prediction module were refined by a CNN-based strategy to obtain the final locations of future GA growth. The 10 scenarios were designed to evaluate the prediction accuracy of our proposed model. The 1-6th scenarios demonstrated the importance of the prior information similarity, the 7-8th scenarios verified the effect of time factors and refinement methods respectively and the 9th scenario compared the prediction results between those using a single follow-up visit for training and using 2 sequential follow-up visits for training. The 10th scenario showed the model generalization performance across regions. The average dice indexes (DI) of the predicted GA regions in the 1-6th scenarios are 0.86, 0.89, 0.89, 0.92 and 0.88, 0.90, respectively. By integrating time factors to the BiLSTM models, the prediction accuracy was improved by almost 10%. The CNN-based refinement strategy can remove the wrong GA regions effectively to preserve the actual GA regions and improve the prediction accuracy further. The prediction results based on 2 sequential follow-up visits showed higher correlations than that based on single follow-up visit. The proposed model presented a good generalization performance while training patients and testing patients were from different regions. Experimental results demonstrated the importance of prior information to the prediction accuracy. We demonstrate the feasibility of creating a model for disease prediction.
自动化预测地理萎缩（GA）病变生长可以帮助眼科医生了解 GA 的进展情况，并评估当前治疗的效果和疾病的预后。我们开发了一种用于识别未来 GA 生长位置的综合时间自适应预测模型。所提出的模型由基于双向长短时记忆（BiLSTM）网络的预测模块和基于卷积神经网络（CNN）的细化模块组成。考虑到连续随访访问之间时间间隔的不连续性，我们将时间因素集成到基于 BiLSTM 的预测模块中，以便方便地控制时间属性。然后，使用基于 CNN 的策略细化预测模块的结果，以获得未来 GA 生长的最终位置。设计了 10 种情况来评估我们提出的模型的预测准确性。前 6 种情况展示了先验信息相似性的重要性，第 7-8 种情况分别验证了时间因素和细化方法的效果，第 9 种情况比较了使用单个随访访问进行训练和使用 2 个连续随访访问进行训练的预测结果。第 10 种情况展示了模型在不同区域的泛化性能。前 6 种情况下预测 GA 区域的平均骰子指数（DI）分别为 0.86、0.89、0.89、0.92 和 0.88、0.90。通过将时间因素集成到 BiLSTM 模型中，预测精度提高了近 10%。基于 CNN 的细化策略可以有效地去除错误的 GA 区域，保留实际的 GA 区域，进一步提高预测精度。基于 2 个连续随访访问的预测结果比基于单个随访访问的预测结果具有更高的相关性。该模型在训练患者和测试患者来自不同区域时表现出良好的泛化性能。实验结果表明了先验信息对预测精度的重要性。我们展示了为疾病预测创建模型的可行性。
