M.S. AAI Capstone Chronicles 2024

CNN Lung Disease Classification

4

Approaching this project as a multi-label classification problem is inherently more challenging than single-label classification. In single-label classification, each instance is associated with only one label, whereas in multi-label classification, an observation can belong to multiple classes simultaneously (Bogatinovski et. al., 2021). For example, we noted a higher number of Infiltration with Effusion cases, and a higher number of Atelectasis with Effusion cases, compared to other co-occurrences. The distribution of these co-occurrences and the number of single label versus multiple labels are shown in the following figures:

During the initial data preprocessing phase, we transformed the raw data through encoding categorical variables and label binarization for multi-label targets. Several variables were included in this dataset, notably the Image Index, Finding Label, Image Width and Height, Image Pixel Spacing, and Image Path. Additional variables such as patient age and gender were also visualized for their distributions. We observed a nearly equal distribution of male and female patients, which ensures our models are sufficiently exposed to X-ray images from both male and female patients during training,

162