But false positives were common in newly reported study

An artificial intelligence (AI) tool trained to identify malignant lesions on CT scans could potentially detect signs of lung cancer a year earlier than current detection methods, researchers reported.

The AI program was shown to be 97% effective for detecting malignant lung nodules in a newly reported study, but false positive readings were also common in the analyzed scans.

The study was presented this week at the virtual annual meeting of the European Respiratory Society – ERS International Congress 2021, in a session entitled “Clinical and Biological Developments in Lung Cancer.”

The session also included new data from a pilot study examining breath analysis as a noninvasive biomarker for lung cancer detection.

The AI study was conducted by the French National Institute for Research in Digital Science and Technology (Inria) at the Universite Cote d’Azur, and preliminary findings were presented by Inria researcher Herve Delingette, PhD.

“The objective of this work was to evaluate a deep-learning system to detect lung nodules from low-dose CT in the context of lung cancer screening,” Delingette said. “We evaluated the performance of this system on an independent dataset and more specifically we were interested in evaluating the ability to detect malignant lesions one year prior to diagnosis by a radiologist.”

The deep learning system was trained on a subset of the public Lung Imaging Database Consortium (LIDC-IDRI) dataset. The learning portion of the study included 888 CT scans, and from this data 2,281 nodules were identified.

The training was conducted using 1,186 nodules characterized as benign or malignant by a consensus of at least 3 (out of 4) radiologists solely based on radiological criteria and the AI program was trained to identify suspicious growths.

The system was then tested on 1,179 patients enrolled in the National Lung Screening Trial study with scanning data available at different time points during three years of follow up.

A total of 177 patients in the cohort were diagnosed with lung cancer by biopsy after their final year in the trial.

Among the 177 malignant nodules, the AI program identified 172 regions of interest within 3 cm of their ground truth location, for a sensitivity of 97%.

When the program was tested on scans taken a year before a lung cancer diagnosis was made, 152 suspicious nodules were identified out of 157 visible nodules that were determined to be malignant.

All 5 undetected lesions were located in the center of the chest, next to the mediastinum, where tumor detection presents a challenge.

“After applying the deep learning system on this test set, we found that it was possible to detect 68% of benign nodules and 97% of malignant nodules,” Delingette said. “Remarkably the AI system was also able to detect 97% of malignant nodules that were visible one year before diagnosis.”

The AI tool also detected, on average, 12 false positive nodules per scan. Delingette said the research team is working to address the issue of false positives.

The breathomics study involved 20 treatment-naive patients in India with confirmed lung cancer (70% male, mean age = 55) and 20 healthy controls (75% male, mean age = 38). The lung cancer patients included 14 with adenocarcinomas and 6 with squamous cell carcinoma. Nine patients had stage III disease and 11 had stage IV disease.

“The exhaled air received from the respiratory tract consist of VOCs, proteins, and RNAs that may be unique to lung cancer morphology,” the researchers noted in their abstract.

Researchers collected breath samples in special bags and analyzed them using a hand held E-Nose device (Cyranose-320). Exhaled Breath Condensate (EBC) was collected and microRNA profiling was performed. Differential expression of microRNAs was analyzed using unpaired t-testing.

Linear discriminate analysis successfully separated breath prints from the lung cancer patients and healthy controls with 85% accuracy (P<0.05), and tumorigenic microRNAs were found to be regulated and anti-tumorigenic microRNAs were downregulated in the EBC samples after the researchers applied multiple normalization algorithms and significant thresholds of P<0.05.

Presenting researcher Bijay Pattnaik of New Delhi, India’s All India Institute of Medical Sciences noted that the exhaled breath analysis accurately differentiated lung cancer patients from healthy controls in the pilot study, suggesting that breathomics could potentially be a non-invasive tool for identifying early disease.

  1. An artificial intelligence tool trained to identify malignant lesions on CT scans could potentially detect signs of lung cancer a year earlier than current detection methods, researchers say.

  2. In a separate pilot study, exhaled breath analysis was found to accurately differentiate lung cancer patients from healthy controls.

Salynn Boyles, Contributing Writer, BreakingMED™

The AI research was conducted by the French national research institute Inria, the University Cote d’Azur and the French AI software company Therapixel. The work was supported by the French National Research Agency and 31A Cote d’Azur Investments. Herve Delingette is an employee of Inria.

Cat ID: 24

Topic ID: 78,24,730,24,192,195,65,925,205