In digital pathology tasks, transformers have achieved state-of-the-art results, surpassing convolutional neural networks (CNNs). However, transformers are usually complex and resource intensive. In this study, we developed a novel and efficient digital pathology classifier called DPSeq, to predict cancer biomarkers through fine-tuning a sequencer architecture integrating horizon and vertical bidirectional long short-term memory (BiLSTM) networks. Using hematoxylin and eosin (H&E)-stained histopathological images of colorectal cancer (CRC) from two international datasets: The Cancer Genome Atlas (TCGA) and Molecular and Cellular Oncology (MCO), the predictive performance of DPSeq was evaluated in series of experiments. DPSeq demonstrated exceptional performance for predicting key biomarkers in CRC (MSI status, Hypermutation, CIMP status, BRAF mutation, TP53 mutation and chromosomal instability [CING]), outperforming most published state-of-the-art classifiers in a within-cohort internal validation and a cross-cohort external validation. Additionally, under the same experimental conditions using the same set of training and testing datasets, DPSeq surpassed 4 CNN (ResNet18, ResNet50, MobileNetV2, and EfficientNet) and 2 transformer (ViT and Swin-T) models, achieving the highest AUROC and AUPRC values in predicting MSI status, BRAF mutation, and CIMP status. Furthermore, DPSeq required less time for both training and prediction due to its simple architecture. Therefore, DPSeq appears to be the preferred choice over transformer and CNN models for predicting cancer biomarkers.

翻译：在数字病理学任务中，Transformer模型已取得超越卷积神经网络（CNN）的最优结果。然而，Transformer通常结构复杂且资源消耗大。本研究开发了一种名为DPSeq的新型高效数字病理分类器，通过微调整合水平与垂直双向长短期记忆网络（BiLSTM）的序列器架构，用于预测癌症生物标志物。基于两个国际数据集——癌症基因组图谱（TCGA）和分子与细胞肿瘤学（MCO）中经苏木精-伊红（H&E）染色的结直肠癌（CRC）组织病理学图像，通过系列实验评估了DPSeq的预测性能。在结直肠癌关键生物标志物（MSI状态、超突变、CIMP状态、BRAF突变、TP53突变及染色体不稳定性[CING]）的预测中，DPSeq展现出卓越性能，在队列内内部验证和跨队列外部验证中均超越多数已发表的最优分类器。此外，在相同实验条件下使用同一训练集与测试集时，DPSeq在预测MSI状态、BRAF突变和CIMP状态时，AUROC与AUPRC值均超越4个CNN模型（ResNet18、ResNet50、MobileNetV2、EfficientNet）和2个Transformer模型（ViT、Swin-T）。由于DPSeq架构简洁，其训练与预测耗时更少。因此，在预测癌症生物标志物时，DPSeq相较于Transformer和CNN模型更具优选性。