Fig. 4: A deep learning model DMAN evaluating mitophagic intermediates by analyzing microscopic images.

A Schematic diagram illustrates the operation of the deep learning model DMAN. First, the training sets were established (left). Images of Mcy3-treated HT22 cells with induced mitophagy were collected. Image patches containing colocalized Mcy3 and GFP-LC3 were used as the training set for mitophagosomes. Image patches with colocalized Mcy3 and Mitotracker served as the training set for mitochondria, while image patches containing colocalized Mcy3 and Lysotracker were used as the training set for mitolysosomes. Mcy3 fluorescence signals from these fluorescence image patches were input into the model. During the testing phase, the outputs include the proportions of mitochondria within different mitophagic intermediates and their pseudo-colored distribution maps in live cells. B Illustration of the dual-branch Multi-scale Attention ResNet (DMAN) model. DMAN is built upon the incorporation of the Residual Network (ResNet) and the Convolutional Block Attention Module (CBAM). During the training phase, DMAN employs a multi-scale approach to extract both fine-grained fluorescence intensity features and coarse-grained morphology features. For fluorescence intensity features, image patches of 32 × 32 pixels are decomposed into three RGB channels. For morphological features, single luminance channels from image patches of 128 × 128 pixels are utilized. The DMAN was trained with the image patches co-labeled with Mcy3, GFP-LC3 and Mitotracker Green as training samples. In the inference phase, a sliding window approach is applied to traverse each pixel in the fluorescence image. The classification of each pixel is determined by the attention scores from both a large window and a small window centered around that pixel. C ROC curves and AUC values in the ten-fold cross-validation. D Confusion matrices for the model’s predictions of the three classes in the ten-fold cross-validation.