Explainable AI for rare blood disorders

Tue, 01 Oct 2019 00:00:00 +0000

Rare anaemias such as hereditary spherocytosis or sickle cell disease are diagnosed from blood samples in which only a fraction of the red blood cells is visibly altered. Labelling every cell is laborious and subjective, specialised genetic tests are expensive and unavailable in most labs, and for rare diseases there are simply very few patients to learn from. Clinicians also need to see why a model reaches a diagnosis before they can trust it. Together with Carsten Marr’s group at Helmholtz Munich and haematology partners (Anna Bogdanova, University of Zurich), we developed explainable methods that diagnose at the level of the patient while pointing to the individual cells that matter.

What we did

Attention-based multiple instance learning. Each blood sample is treated as a bag of single red blood cells detected with an R-CNN; an attention-based MIL model classifies the patient into one of four blood cell disorders, and the attention weights show how much each cell contributed — improving both accuracy and interpretability for the haematologist ( ).
Percoll density gradients as an affordable biomarker. Centrifuging a patient’s blood through a Percoll gradient separates red cells by density. Combining CNN features with Fourier spectral features of the gradient images classified hereditary haemolytic anaemias with an average F1-score of 88% ( ), and graph convolutional networks on the gradient profiles predicted sickle cell disease severity ( ).
Robustness across labs. A cross-domain adapted autoencoder extracts white blood cell features in an unsupervised way, so that classifiers trained on one dataset transfer to smears from other sites, stains and microscopes ( ).
Anomaly-aware MIL for very rare disorders. By using healthy cells from negative bags as a reference, an interpretable pooling method raises the contribution of anomalous cells and outperforms attention-based MIL on rare anaemia classification with few training samples ( , with Salome Kazeminia).

Our role

Co-supervision of Ario Sadafi’s PhD (TU Munich and Helmholtz Munich, with Carsten Marr and Nassir Navab), started at TU Munich CAMP and continued at Helmholtz AI, and contributions to the work of Salome Kazeminia and Raheleh Salehi in Carsten Marr’s group.

Perspective

The project showed that patient-level labels are enough to train cell-level explanations, a principle we now apply to weakly supervised learning in computational pathology and other rare-disease settings.

Multiple Instance Learning |

Explainable AI for rare blood disorders

What we did

Our role

Perspective