A comparative study of tumor detection models trained on coronal versus sagittal versus axial PET imaging slices

Introduction: Extracting PET imaging biomarkers such as the total metabolic tumor volume (TMTV) from segmented tumors has the potential to improve assessment of disease and prediction of patient outcome. Due to the time-consuming nature of manual tumor segmentation, however, such biomarkers are not quantified routinely. Automated detection of tumor bounding boxes can be an important step towards biomarker quantification, as automatic segmentation can be carried out inside the boxes more easily than on whole-body images.

In this work, we focus on primary mediastinal large B-cell lymphoma (PMBCL), an lymphoma subtype. We train a deep detection network FasterRCNN to detect tumors on the 2D slices of PMBCL PET images. We train 3 detection models on coronal, sagittal and axial slices and compare their performance on the test set.

Methods: Our PMBCL PET dataset (n=126) was segmented by nuclear medicine physicians, following the development of consensus procedures. These cases were first 60:20:20% split into train, validation and test sets. From these 3D images, 3 datasets were created: consisting of coronal, sagittal and axial slices. The class imbalance between foreground (FG) slices (slices intercepting tumors) and background (BG) slices (slices not intercepting tumors) was 9:91%, 11:89% and 10:90% for the coronal, sagittal and axial datasets respectively across all train, validation and test sets. The FG slices were augmented using translation, rotation and scaling to obtain a balanced set between FG and BG slices.

For the detection task, FG slices were annotated with tight bounding boxes (tumor) around the tumors. For BG slices, center of the box (denoting BG) was chosen to be the center of the slice and box dimension was set nearly equal to slice dimension.

We used an ImageNet-pretrained FasterRCNN model and employed stochastic gradient descent to optimize the object detection loss function (weighted sum of cross-entropy loss for object detection and L1 loss for box offset). The 3 detection models (namely coronal, sagittal and axial models) were trained for 100, 110 and 80 epochs respectively. For accepting a prediction of these detection models, we used an intersection over union (IOU) threshold=0.5 between the ground truth and predicted boxes. The performance of the 3 detection models was compared using detection accuracy (AC) and mean average precision (mAP) metrics on the test sets

Results: We achieved AC=52% and mAP=0.46, AC=53% and mAP=0.50, and AC=70% and mAP=0.66 for coronal, sagittal and axial models respectively on test set. The axial model outperformed the other two models on both AC and mAP metrics. The reasons for this can be multifold (i) there were about 1.3 times more slices in the axial set as compared to coronal/sagittal sets, owing to the larger size of 3D PET images along the axial direction. Hence, the axial model was trained on more single slice samples; (ii) As each axial slice intercepts a smaller cross-section of the body as compared to coronal/sagittal slices, on average the axial FG slices intercepted less number of tumors (1.32tumors/FG slice) as compared to coronal FG slices (2.11tumors/FG slice) or sagittal FG slices (1.78tumors/FG slice), making it easier for detection model to localize them on axial slice. The histograms for number of tumors on FG slices are given in Fig 1

Conclusions: In this work, we study the importance of selecting the slice orientation for creating datasets for tumor detection. Choosing the orientation such that the slice dataset contains smaller number of objects per image can help the model learn better localization skills. We will also be exploring other techniques, including 2.5D multi-slice input methods, as well as collective utilization of coronal/sagittal/axial slices for improved training of a given model.

• Download figure
• Open in new tab
• Download powerpoint