A cascaded deep network for automated tumor detection and segmentation in clinical PET imaging of diffuse large B-cell lymphoma

TL;DR

This work tackles automated detection and segmentation of Diffuse Large B-Cell Lymphoma (DLBCL) in whole-body 18F-FDG PET images by introducing a three-step cascaded network that sequentially classifies slices, detects tumors, and segments ROIs. Each module is trained independently (slice classifier: ResNet152; tumor detector: Faster R-CNN with ResNet50+FPn; tumor segmentor: 2D U-Net), collectively delivering improved 3D segmentation accuracy with a 3D Dice score of 78.1% on the DLBCL test set, compared with 58.9% from a single-end-to-end U-Net. The approach yields strong detection metrics (slice classifier AUC 0.93; detector mAP 0.69; accuracy 81%) and demonstrates enhanced performance by focusing segmentation within ROIs, suggesting clinical potential for rapid TMTV estimation and personalized therapy planning.

Abstract

Accurate detection and segmentation of diffuse large B-cell lymphoma (DLBCL) from PET images has important implications for estimation of total metabolic tumor volume, radiomics analysis, surgical intervention and radiotherapy. Manual segmentation of tumors in whole-body PET images is time-consuming, labor-intensive and operator-dependent. In this work, we develop and validate a fast and efficient three-step cascaded deep learning model for automated detection and segmentation of DLBCL tumors from PET images. As compared to a single end-to-end network for segmentation of tumors in whole-body PET images, our three-step model is more effective (improves 3D Dice score from 58.9% to 78.1%) since each of its specialized modules, namely the slice classifier, the tumor detector and the tumor segmentor, can be trained independently to a high degree of skill to carry out a specific task, rather than a single network with suboptimal performance on overall segmentation.

Figures (3)

• Figure 1: The proposed cascaded 3-step deep network for PET tumor segmentation. Module-1 (Slice classifier) is a ResNet152 based binary classification network trained on all axial PET slices. Module-2 (Tumor detector) is a Faster R-CNN based object detection network trained only on the axial foreground PET slices. Module-3 (Tumor segmentor) is a 2D U-Net network trained on ROIs extracted from the ground truths of Module-2.
• Figure 2: Performance of the slice classifier module on the DLBCL test set. (a) Various classification metrics as a function of threshold. (b) The optimal threshold = 0.16 was chosen using the maximum of the geometric mean of specificity and sensitivity. (c) Receiver operating characteristic (ROC) curve, with the area under curve (AUC) = 0.93. (d) Precision-recall (PR) with average precision = 0.72.
• Figure 3: Performance of the tumor detector and segmentor modules: (left column) (a)-(c) show 3 representative DLBCL PET images in the coronal view. (middle column) (d)-(f) show the corresponding selected axial slices (shown as white horizontal lines in (a)-(c)) with the predicted bounding boxes (ROIs) around the tumors (shown in red). The average detection accuracy on the DLBCL test set was 81% and average mAP was 0.69. (right column) (g)-(i) show the corresponding ground truth (via physicians, shown in green) and predicted (via our implemented 2D U-Net, shown in red) segmentation contours in 2D, with the average 2D Dice score of 77.9% $\pm$ 13.2% across all the slices in the DLBCL test set obtained as output of the detection module. The 2D segmentation contours were aggregated to obtain an average 3D Dice score of 78.1% $\pm$ 8.6% across all test patients.