Department of Thoracic Surgery, Sichuan Academy of Medical Sciences and Sichuan Provincial People's Hospital, University of Electronic Science and Technology of China, Chengdu, China.
Institute of Thoracic Oncology and Department of Thoracic Surgery, West China Hospital, Sichuan University, Chengdu, China.
Br J Cancer. 2024 Dec;131(11):1833-1845. doi: 10.1038/s41416-024-02856-8. Epub 2024 Oct 25.
Non-small cell lung cancer (NSCLC) is one of the leading causes of cancer mortality worldwide. Immune checkpoint inhibitors (ICIs) have emerged as a crucial treatment option for patients with advanced NSCLC. However, only a subset of patients experience clinical benefit from ICIs. Therefore, identifying biomarkers that can predict response to ICIs is imperative for optimising patient selection.
Hematoxylin and eosin (H&E) images of NSCLC patients were obtained from the local cohort (n = 106) and The Cancer Genome Atlas (TCGA) (n = 899). We developed an ICI-related pathological prognostic signature (ir-PPS) based on H&E stained histopathology images to predict prognosis in NSCLC patients treated with ICIs using deep learning. To accomplish this, we employed a modified ResNet model (ResNet18-PG), a widely-used deep learning architecture well-known for its effectiveness in handling complex image recognition tasks. Our modifications include a progressive growing strategy to improve the stability of model training and the use of the AdamW optimiser, which enhances the optimisation process by adjusting the learning rate based on training dynamics.
The deep learning model, ResNet18-PG, achieved an area under the receiver operating characteristic curve (AUC) of 0.918 and a recall of 0.995 on the local cohort. The ir-PPS effectively risk-stratified NSCLC patients. Patients in the low-risk group (n = 40) had significantly improved progression-free survival (PFS) after ICI treatment compared to those in the high-risk group (n = 66, log-rank P = 0.004, hazard ratio (HR) = 3.65, 95%CI: 1.75-7.60). The ir-PPS demonstrated good discriminatory power for predicting 6-month PFS (AUC = 0.750), 12-month PFS (AUC = 0.677), and 18-month PFS (AUC = 0.662). The low-risk group exhibited increased expression of immune checkpoint molecules, cytotoxicity-related genes, an elevated abundance of tumour-infiltrating lymphocytes, and enhanced activity in immune stimulatory pathways.
The ir-PPS signature derived from H&E images using deep learning could predict ICIs prognosis in NSCLC patients. The ir-PPS provides a novel imaging biomarker that may help select optimal candidates for ICIs therapy in NSCLC.
非小细胞肺癌（NSCLC）是全球癌症死亡的主要原因之一。免疫检查点抑制剂（ICIs）已成为晚期 NSCLC 患者的重要治疗选择。然而，只有一部分患者从 ICI 中获得临床获益。因此，确定能够预测 ICI 反应的生物标志物对于优化患者选择至关重要。
从本地队列（n=106）和癌症基因组图谱（TCGA）（n=899）获得 NSCLC 患者的苏木精和伊红（H&E）图像。我们基于 H&E 染色组织病理学图像开发了一种与 ICI 相关的病理预后签名（ir-PPS），以使用深度学习预测接受 ICI 治疗的 NSCLC 患者的预后。为了实现这一目标，我们采用了一种经过修改的 ResNet 模型（ResNet18-PG），这是一种广泛使用的深度学习架构，以其在处理复杂图像识别任务方面的有效性而闻名。我们的修改包括渐进式增长策略，以提高模型训练的稳定性和使用 AdamW 优化器，该优化器通过根据训练动态调整学习率来增强优化过程。
深度学习模型 ResNet18-PG 在本地队列上获得了 0.918 的接收器工作特征曲线下面积（AUC）和 0.995 的召回率。ir-PPS 有效地对 NSCLC 患者进行了风险分层。与高危组（n=66）相比，接受 ICI 治疗后低危组（n=40）的无进展生存期（PFS）显著改善（对数秩 P=0.004，风险比（HR）=3.65，95%CI：1.75-7.60）。ir-PPS 对预测 6 个月 PFS（AUC=0.750）、12 个月 PFS（AUC=0.677）和 18 个月 PFS（AUC=0.662）具有良好的区分能力。低危组表现出免疫检查点分子、细胞毒性相关基因的高表达、肿瘤浸润淋巴细胞的丰富程度增加以及免疫刺激途径的活性增强。
使用深度学习从 H&E 图像中提取的 ir-PPS 特征可以预测 NSCLC 患者的 ICI 预后。ir-PPS 提供了一种新的成像生物标志物，可能有助于选择 NSCLC 患者接受 ICI 治疗的最佳候选者。
