Machine Learning Intern - Learning from radiology reports to detect cancer on CT Raidium

Raidium develops a radiological foundation model as the “GPT” of radiology. This new generation of AI will enable the building of an imaging biomarker factory for both clinical practice and research, tackling the complexity of precision medicine

Oncology accounts for a large share of radiologists’ workload, for diagnosis to treatment. At Raidium, we built OncopilotDetect, a universal lesion detector that finds lesions, benign or malignant, across all organs on chest, abdomen and pelvis CT. While localisations and measurements of the lesions are already helpful, a significant part of a radiologist’s time is spent on assessing their malignancy and other relevant characteristics.

Training a model to characterise lesions would usually require lesion-level labels. These are expensive, because experts have to exhaustively describe each lesion of many CT exams, which is not done during routine reading of exams, and is very time-consuming. Radiology reports, on the other hand, exist for every exam series, but leveraging them is an open problem, as reports often describe information contained in one of the related CT exams of a series, without the same level of information (location, characterisation) for each lesion, and usually aggregate less important lesions’ descriptions into a common sentence.

Recent work such as R-Super (Bassi et al., MICCAI 2025) shows that partial information on the count, size and location of some lesions found in radiological reports can supervise lesion segmentation at scale. The goal of this internship is to learn lesion-level characterisation from report-derived labels, using OncopilotDetect’s detections as the lesion candidates.

In this internship, you will:

• Structure a training dataset from existing data sources (CT exams, radiology reports, outputs of our detection models), with report-derived labels: number of lesions per class, size, location and uncertainty.

• Design a weakly supervised method that learns to characterise each lesion from these labels, using the size and location constraints given by the report.

• Handle the specifics of radiology reports: hedged wording (“probable hemangioma”, “indeterminate”), findings left unreported when a more severe one dominates, and label noise.

• Learn how radiologists characterise lesions (guideline categories such as Bosniak or LI-RADS) and work with them to build a lesion-level test set with a reliable reference standard.

• Benchmark your approach against fully supervised baselines, with the goal of a publication at a venue such as MICCAI.