Machine learning device for quickly identifying tissue malignancy during live surgery.

When a patient presents with potentially malignant pulmonary tissue via pre-operative imaging, surgery and early removal of the cancerous tissue is their best avenue for treatment. However, when proceeding with the operation, the surgeon must remove a diagnostic resection of the lesion and wait for the pathology department to confirm the presence of cancer, a process that adds up to 35 minutes to the procedure. Recent developments in endomicroscopy provide cellular resolution imaging that allows for the classification of tissues without the need for lengthy diagnostic tests, but the interpretation of such visual data requires extensive training for operators. 

The inventors developed a machine learning methos that assesses the potential malignancy of tumor cells in real-time in the operating room. The device only takes a few seconds to perform classification and evaluates the 3D structure of the tumor. Furthermore, it also indicates any potential sampling errors to further lower the risk of user error.

Near-infrared needle-based confocal laser endomicroscopy (NIR-nCLE) is an imaging method providing cellular level resolution of potential tumors. The inventors used NIR-nCLE combined with an NIR contrast imaging agent to build a high-fidelity training set with over 1500 videos and trained a convolutional neural network to identify malignant tissues. 

• Predicts malignancy of tissue with >93.5% accuracy
• Performs prediction on single-cell nCLE imaging in real time
• Informs operator if tissue sampling is adequate
• Device can be tailored for cloud-based remote lesion assessment