Detection of malignant lung tumors using stimulated Raman histology and convolutional neural networks

Patients with lung tumors often receive their histopathological diagnosis intraoperatively, based on hematoxylin and eosin-stained frozen sections. However, this approach is time and labour-intensive. Intraoperative stimulated Raman histology (SRH) may bypass traditional histopathologic processing as it leverages stimulated Raman scattering (SRS) of photons at molecules in fresh tissue samples to generate histologic images. Automated image analysis using convolutional neural networks (CNNs) could further accelerate intraoperative histopathological diagnosis. This study aimed to investigate CNN-based detection of lung cancer and the ability to distinguish between histologic subtypes, primary lung tumors, and pulmonary metastasis.

A total of 459 fresh frozen tissue samples were obtained from 133 patients undergoing lung resection for intrapulmonary malignancies. SRS and SRH images were acquired, images were annotated, and three CNNs were trained and evaluated on both SRS and SRH images.

When distinguishing between intrapulmonary malignancy and normal lung tissue, the three different CNNs VGG19 achieved a balanced accuracy of 0.89 (0.95 on SRH images), ResNet50 achieved a balanced accuracy of 0.87 (0.89 on SRH images), and Inception-ResNet-v2 achieved a balanced accuracy of 0.91 (0.94 on SRH images). On SRS images, Inception-ResNet-v2 (0.91) showed the best results, followed by VGG19 (0.89). Compared to SRS images, the SRH images show higher balanced accuracy. Neither a distinction between primary lung cancer and metastases nor between squamous cell carcinoma and adenocarcinoma was achieved.

The results of this study demonstrate the ability of CNNs to identify malignant tumors of the lung on SRS and SRH images. A distinction between different World Health Organization (WHO) subtypes of primary lung cancer and between primary lung cancer and metastases was inaccurate in our dataset.