The biostatistics R&D center of a global pharmaceutical laboratory noted the ineffectiveness of existing methods for classifying cancer by imaging. We have shown that deep learning can surpass these methods even on small samples while remaining interpretable by pathologists.
Cancer detection via deep learning
The biostatistics R&D center of a global pharmaceutical laboratory noted the ineffectiveness of existing methods for classifying cancer by imaging. We have shown that deep learning can surpass these methods even on small samples while remaining interpretable by pathologists.
Problem
Solution
What we built
We deployed 1 Data Scientist and 1 Data Engineer in Scrum mode to design a POC for the classification of medical images by Deep Learning, which can be validated by pathologists.
Step 1 — Data preparation. Creation and cleaning of the data set, normalization and data augmentation on the fly to compensate for the reduced sample size and maximize the robustness of the model.
Step 2 — Deep Learning modeling. Development of a convolutional neural network (CNN) based on a ResNet architecture, with transfer learning to take advantage of pre-trained models and progressive resizing to optimize learning on a small dataset.
Step 3 — Interpretability. Implementation of a complete explainability system: confusion matrix, performance metrics (accuracy, precision, recall) and heatmaps showing the areas of the image that influenced the classification. Pathologists could verify that the model was based on the right clinical characteristics.
Step 4 — Validation by experts. Presentation of the POC to the R&D department and pathologists. Validation of the results and clinical characteristics identified by the model.
Projects in the same category
