Two-Stage Deep Learning Model for Automated Segmentation and Classification of Splenomegaly

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Splenomegaly is a frequent finding in cross-sectional imaging, presenting diverse diagnostic challenges.
• Differentiating causes of splenomegaly, such as cirrhotic portal hypertension and lymphoma, is crucial for patient management.

Purpose of the Study:

• To develop and evaluate a deep learning model for automatic spleen segmentation.
• To assess the model's capability in identifying the cause of splenomegaly using computed tomography (CT) images.
• To compare classification performance between models trained on whole abdominal CT scans versus spleen segmentation masks.

Main Methods:

• A retrospective study involving 149 patients with splenomegaly (77 cirrhotic portal hypertension, 72 lymphoma) undergoing CT scans.
• Spleen segmentation was performed using a modified U-Net architecture.
• Classification of splenomegaly causes was achieved using a 3D DenseNet, evaluated on whole abdominal CT and spleen masks.
• Performance metrics included area under the receiver operating characteristic curve (AUC), accuracy (ACC), sensitivity (SEN), and specificity (SPE).

Main Results:

• The deep learning model achieved high performance in differentiating lymphoma from liver cirrhosis when trained on whole abdominal CT volumes (AUC = 0.88, ACC = 0.88).
• Training the model using only the spleen segmentation mask resulted in decreased performance (AUC = 0.81, ACC = 0.76).
• Occlusion sensitivity maps highlighted important regions for prediction in whole abdominal CT images.

Conclusions:

• The developed deep learning model demonstrates efficacy in segmenting splenomegaly and identifying its underlying cause.
• Utilizing whole abdominal CT scans for training outperformed models trained solely on spleen segmentation masks.
• The model's performance suggests potential for broader applications in diagnosing various causes of splenomegaly.