Radiology: Histopathological value of a deep learning survival prediction model for chest CT in lung adenocarcinoma patients

At present, various methods have been tried to predict the prognosis of lung cancer patients in clinical practice, and imaging has played a key role
in these attempts.
The size of the tumor and the presence of ground-glass lesions have been repeatedly reported as radiographic prognostic factors
for lung adenocarcinoma.
In addition, various studies of radiomic or textured features have shown that chest CT scans provide potential prognostic information
in addition to anatomical information.
Recently, a CT-based preoperative deep learning (DL) prediction model has been proposed to estimate disease-free survival in lung adenocarcinoma patients
.
The DL model demonstrated discriminating performance comparable to clinical class T and functioned as an independent survival predictor after adjusting for age, sex, lobe position, and clinical class T.

However, a major challenge with survival prediction models is that clinicians lack the confidence
to make the corresponding diagnosis.
Especially with DL models, the inference process is difficult to explain and lacks a logical basis
.
This black-box nature of DL makes clinicians reluctant to use these models
in clinical work.

Recently, a study published in the journal Radiology evaluated the association between features extracted from DL survival prediction models and histopathological risk factors by using unsupervised clustering studies, and analyzed the value of these correlations through a series of regression analyses, providing a reference for
clinical accurate assessment of the prognosis of lung cancer patients.

This retrospective study collected data from patients with lung adenocarcinoma who underwent radical resection between January 2016 and September 2020 from a tertiary care centre, each without neoadjuvant therapy
.
Seven histopathological risk factors for resectable adenocarcinoma were documented: aggressive adenocarcinoma subtypes; mediastinal nodule metastasis (pN2); whether there is lymphatic or venous invasion; Visceral pleural invasion (VPI); and EGFR mutation state
.
Unsupervised clustering was performed using 80 DL model-driven CT features, and associations
between patient clustering and histopathological features were analyzed.
Multivariate regression analysis was performed to investigate the added value of DL model output to semantic CT features (clinical class T and radiological nodule type [i.
e.
, solid or subsolid]) in histopathology.

A total of 1667 patients were evaluated (median age, 64 years [IQR, 57-71 years]; 975 women).

In addition to the EGFR mutation state (P=0.
30 for cohort 3), unsupervised patient cohorts 3 and 4 were associated with all histopathological risk factors (P<0.
01).

。 After multivariate adjustment, model outputs were associated with aggressive adenocarcinoma subtype (probability [OR], 1.
03; 95% CI: 1.
002, 1.
05; P = .
03), venous invasion (OR, 1.
03; 95% CI: 1.
004, 1.
06; P = .
02), and VPI (OR, 1.
08; 95% CI: 1.
06, 1.
10; P < .
001), but not
semantic CT features.

The box plot shows the deep learning (DL) model output based on the presence of
(A) adenocarcinoma subtype, (B) mediastinal nodule metastases, (C) lymphatic invasion, and (D) venous invasion.
The output of the DL model is clearly correlated with all seven histopathological risk factors, supporting the biology of the DL model and the reliability of its results in predicting patient survival

This study shows that the CT-based preoperative deep learning prediction model can extract the CT features of the histopathological and metabolic features of lung adenocarcinoma, which is conducive to the clinical prediction of postoperative survival of
lung cancer patients.

Original source:

Ju G Nam,Samina Park,Chang Min Park,et al.
Histopathologic Basis for a Chest CT Deep Learning Survival Prediction Model in Patients with Lung Adenocarcinoma.
DOI:10.
1148/radiol.
213262