Trainee Project
Title:
Multi-modal classification of in vivo spectroscopic data for the differential diagnosis of skin cancers
Dates:
2026/02/18 - 2026/08/30
Student:
Supervisor(s):
Description:
The in vivo characterisation of biological tissues is a fundamental challenge for the rapid and non-invasive diagnosis and monitoring of skin cancers. In this context, optical spectro-imaging methods play a key role. The SpectroLive device has enabled the clinical acquisition of a database containing autofluorescence (AF) and diffuse reflectance (DR) spectra of various skin lesions labelled as cancerous, pre-cancerous or healthy. In addition, high-resolution (HR) images of the histological sections on which the anatomopathological examinations that enabled this categorisation were performed are also available.
The objective of this internship will be to implement and test machine learning methods for the classification of different lesions, taking advantage of the three modalities available: AF, DR and HR.
1) Segmentation of the HR images of sections
The segmentation of lesion areas on HR captures of tissue sections would complement the anatomopathological diagnoses already obtained, thereby enabling more precise tissue characterisation (pathology stages, etc.). Existing neural network-based approaches will be adapted to perform this task.
2) Multi-modal classification
The combined use of AF and RD modalities for spectrum classification is an important challenge in the processing of available spectroscopic data. A previous work has highlighted the value of optimal transport distances for spectra clustering. We will now focus on extensions that would allow the simultaneous integration of both modalities, such as vector-valued optimal transport.
The objective of this internship will be to implement and test machine learning methods for the classification of different lesions, taking advantage of the three modalities available: AF, DR and HR.
1) Segmentation of the HR images of sections
The segmentation of lesion areas on HR captures of tissue sections would complement the anatomopathological diagnoses already obtained, thereby enabling more precise tissue characterisation (pathology stages, etc.). Existing neural network-based approaches will be adapted to perform this task.
2) Multi-modal classification
The combined use of AF and RD modalities for spectrum classification is an important challenge in the processing of available spectroscopic data. A previous work has highlighted the value of optimal transport distances for spectra clustering. We will now focus on extensions that would allow the simultaneous integration of both modalities, such as vector-valued optimal transport.
Keywords:
Segmentation, classification, multimodal, optimal transport
Department(s):
| Biology, Signals and Systems in Cancer and Neuroscience |