I once worked with physicians and other co-authors on a research manuscript about deep learning in breast cancer histopathology.
The research question concerned predicting and detecting axillary lymph-node metastasis. It sits at the intersection of whole-slide images, pathology, machine learning and decisions that have real clinical consequences.
My most useful takeaway was not a specific neural-network architecture. It was a reminder that a technically sophisticated model can still answer the wrong question if domain definitions and data lineage are weak.
An image is not just an image
A histopathology slide carries context that a model does not automatically understand:
- how the tissue was sampled and prepared;
- which region a pathologist considers relevant;
- how labels were derived;
- whether multiple slides belong to the same patient;
- what “positive” means for the research question;
- which differences come from biology and which come from scanning or preparation.
Without that context, a model may learn shortcuts. It can appear accurate while relying on properties that will not survive a new hospital, scanner or patient cohort.
Domain experts change the problem definition
Working with physicians changes the conversation from “Can we train a classifier?” to better questions:
- Is the target clinically meaningful?
- Does the label represent ground truth or an imperfect proxy?
- Could information from one patient leak across training and evaluation?
- Which error is more serious in this setting?
- What evidence would be needed before anyone could act on the output?
These are data-quality questions as much as machine-learning questions.
Evaluation must match the unit of use
Medical image datasets often contain many patches or slides for a smaller number of patients. A random image-level split can therefore give an optimistic result if material from one patient appears on both sides.
The evaluation design has to respect the real unit of use: usually the patient, the case, or an independent cohort. Metrics also need clinical interpretation. A single accuracy number cannot explain sensitivity, specificity, class imbalance, calibration or the consequences of an error.
Related public research
The separate BCNB public challenge is a useful example of this broader research area. It provides breast-cancer core-needle biopsy whole-slide images, clinical characteristics and tasks that include prediction of axillary lymph-node metastatic status.
I link it as context, not as the dataset or publication from our collaboration.
Evidence boundary
I have records of the collaboration and manuscript submission, but I did not find a public publication under the working title supplied to me. I am therefore not presenting the manuscript as published or peer reviewed, and I am not publishing private correspondence, patient data, model metrics or unpublished figures.
If a public paper becomes available, this note can be updated with the final citation and verified results.
The lesson I carried into other work
The same principle applies outside medicine.
If a data team cannot explain where a label came from, what one row represents, how train and test records are related, or what an error costs, a more complex model will not repair the foundation.
Start with provenance, definitions, leakage controls and an evaluation that resembles the real decision. The model comes after that.