Tumorous Tissue Classification
In collaboration with the BioRobotics Institute of Scuola Superiore Sant'Anna, we developed a machine learning pipeline for classifying tissue-mimicking materials using a custom robotic probe equipped with force and position sensors (Massari et al., 2019; Bernardo et al., 2025). The long-term goal is to support intraoperative biopsy decisions by distinguishing tumorous from healthy tissue in real time.
The current work uses a custom dataset of five synthetic materials embedded in a silicone vessel to mimic organic tissue properties, providing a controlled testbed before transitioning to real biological samples.
Experimental Setup
The robotic probe integrates force and position sensors to record interaction signals as it scans across the surface of the silicone vessel. Five different materials with distinct mechanical properties are arranged inside, each representing a class to be identified.
Signal preprocessing was critical: raw sensor outputs were highly noisy due to environmental variability (temperature, humidity, probe contact angle). We applied filtering to remove border effects at material transitions, and guided feature engineering by domain literature to extract physically meaningful descriptors characterising each material’s mechanical response.
Classification Pipeline
We evaluated classical machine learning techniques across the engineered feature space. To reduce misclassification at material boundaries, we aggregated predictions over multiple consecutive probe measurements before committing to a label. An ensemble model combining several classifiers provided the best prototype, achieving strong results on the held-out test set.
