Cancer Therapy

For patients with locally advanced non-small cell lung cancer (NSCLC), deep learning networks integrating compute tomography (CT) scans at multiple time points can improve clinical outcome predictions, according to a study published online April 22 in Clinical Cancer Research.

Common cancers worldwide

Lung cancer is one of the most common cancers worldwide; so the highest contributor to cancer death in both the develop and developing worlds. Among these patients, most are diagnose with non small cell lung cancer (NSCLC) and have a 5-year survival rate of only 18%. Despite recent advancements in medicine spurring a large increase in overall cancer survival rates; so this improvement is less consequential in lung cancer, as most symptomatic and diagnosed patients have late-stage disease.

Using single seed-point tumor localization, models were develop using transfer learning of convolutional neural networks (CNN) with recurrent neural networks. Pathologic response validation was perform on dataset B, which include 178 scans from 89 patients treat with chemo radiation and surgery.

The researchers found that survival and cancer-specific outcomes (progression, distant metastases, and local-regional recurrence) were predict by deep-learning models using time series scans. With each additional follow-up scan into the CNN model, model performance was enhanced. Patients were stratified into low and high mortality risk groups by the models, which were significantly associate with overall survival (hazard ratio, 6.16).

Prediction of survival

Tracking tumor evolution for prediction of survival and response after chemotherapy and radiation therapy can be critical to treatment assessment and adaptive treatment planning for improving patient outcomes. Conventionally, clinical parameters are used to determine treatment type and to predict outcome; but this does not take into account phenotypic changes in the tumor.

Medical imaging tracks this evolution of lesions noninvasively; also provides a method for tracking the same region longitudinally through time; so providing additional tumor characteristics beyond those obtain through static images at a single time point. This study demonstrate the impact of deep learning on tumor phenotype; so tracking before and after definitive radiation therapy through pre treatment and CT follow up scans.

Although the input of this model consisted of a single seed point input at the center of the lesion; so without the need for volumetric segmentation our model; which had comparable predictive power compare with tumor volume, acquired through time-consuming manual contours. Noninvasive tracking of the tumor phenotype predicted survival, prognosis, and pathologic response; which can have potential clinical implications on adaptive and personalized therapy.