Leveraging prior knowledge in machine intelligence for improving cancer diagnostic accuracy

As a representation of machine intelligence, machine learning (ML) has made significant strides in emulating the expertise of medical professionals in interpreting medical image characteristics to predict lesion malignancy for medical imaging-based diagnosis of early cancers. The experts’ interpretations of medical images usually incorporate prior knowledge about the tissue pathology of malignancy or benignity and how the images were generated into decision making for the diagnosis task. However, the experts’ incorporation of prior knowledge is mainly in a qualitative manner based on their past training and experience in the field. We hypothesize that quantitative incorporation of prior knowledge by ML via mathematical formulas or computer programming can improve the performances of the experts and current ML algorithms for the diagnosis task. This study aims to test this hypothesis by leveraging prior knowledge in machine intelligence (pkMI) to improve lesion diagnosis in low dose computed tomography (CT) screening for early lung cancer detection. One key piece of prior knowledge relates to the CT’s X-ray energy spectrum, where different energies interact with in vivo tissues inside a lesion and generate variable but reproducible image contrasts, encapsulating tissue biological information in the image contrast variations for the diagnosis task. Another critical piece of prior knowledge involves the dynamic or functional properties of in vivo tissues, such as elasticity, which indicates pathological conditions for the diagnosis task. The pkMI was evaluated using two datasets of indeterminate lesions (IDLs) with their pathological reports as indication of malignancy or benignity. The hypothesis was confirmed, showing a dramatic improvement in diagnosing the IDLs’ malignancy by the area under receiver operating characteristic curve from the 0.70s up to the 0.90s.