Table 1.
Demographic information of the 769 patients.
Fig 1.
(A) LSTM block and (B) Multi-head self-attention.
Fig 2.
(A) The process of identifying representative regions of interest (ROIs) in an image from large whole-slide images (WSIs). (B) Fine-tuning the ResNet50 pretrained on the ImageNet dataset via additional ROI-based training. After fine-tuning, CNN weights of half of the front layers are frozen. This CNN is used as a feature extraction network for DALAN. (C) The patient’s ROIs are randomly sampled with replacement, and then they undergo data augmentation and shuffling before being fed into DALAN’s input.
Fig 3.
The overall structure of DALAN.
The lesion images are compressed into embedding vectors using pretrained CNN and then processed through two Attention-LSTM blocks, which use a multi-head (co-)attention layer and a two-stacked LSTM layer to calculate the significance of each lesion and produce a comprehensive representation of the patient’s lesion data. Cox regression is then performed through an MLP to produce the final hazard prediction. During training, N lesion ROIs are randomly sampled, and data augmentation is performed to improve the model’s performance and robustness.
Fig 4.
Workflow of the simulation study.
(A) MNIST dataset simulation. Simulated survival times T were generated for two ROIs representing the digits "0" and "6," and these survival times were averaged to obtain the final survival time for the sample. (B) Cancer dataset simulation. Likewise, simulated survival times T were generated for multiple ROIs, and the final survival time for the sample was obtained by averaging the survival times of all ROIs.
Table 2.
Performance comparison with competing methods on the simulated dataset in terms of c-index (mean±SD).
Table 3.
Performance comparison with competing methods on TCGA dataset in terms of c-index (mean±SD).
Table 4.
Performance comparison with ablation study on simulated and TCGA dataset in terms of c-index (mean±SD).
Fig 5.
Kaplan-Meier plots according to grade, IDH status, and DALAN.
Fig 6.
Distribution of normalized predicted risk scores according to gender, grade, molecular IDH status, and age.
Table 5.
Statistical analysis of predicted risk scores.