Figure 1
MILLIE approach. (a) Preprocessing and training data generation from a peripheral blood smear. Fields of view acquired with a high-NA objective lens (100×/1.4NA for blood films and 50×/0.55 for BMA) digital microscope are processed. A histogram-based segmentation28 (see “Methods” section) was employed to generate binary masks corresponding to individual cells from the RGB images. These masks are further employed to crop individual patch images around each individual cell from the RGB images. (b) Training with weak labels. The extracted patches are passed through the convolutional neural network. Corresponding convolutional feature vectors are pooled together in one single feature vector (max pooling) followed by fully connected and classification layers. Weights of the model are optimized to predict the sample-level label available from routine clinical examinations. (c) Detecting morphological indicators. Once trained, individual cells can be passed one-by-one through the MILLIE models which classifies them as indicators for the specific disorders MILLIE learned to predict at a sample level.
