Towards enhanced automated tumor detection using background slice annotation methods in clinical PET imaging of lymphoma

Introduction: Detection of lymphoma tumors from PET images is paramount to segmentation. Tumor segmentation is required for calculating quantitative metrics such as the total metabolic tumor volume (TMTV) and other radiomics features with prognostic value for patient outcome in lymphoma. Manual segmentation of tumors in whole-body images is time-consuming; hence not performed routinely. Placing boxes around tumors will help physicians read the scans faster by pointing their attention to the tumors and is a first step towards automated segmentation. The segmentation can be performed more efficiently inside the box.

In this work, we focus on primary mediastinal large B-cell lymphoma (PMBCL) which is an aggressive lymphoma subtype. We train an object detection model Faster RCNN to localize tumors on the axial slices of PMBCL PET images. We propose to improve our model by various methods for annotating the background (BG) axial slices (slices not intercepting tumors).

Methods: Our PMBCL PET dataset (n=126) were annotated by expert nuclear medicine physicians, following development of consensus procedures. We employed 60:20:20 % split of axial slices between training, validation and test sets respectively. All 2D slices were resized to 224×224 pixels and stacked thrice forming 3-channel inputs. The imbalance between foreground (FG) slices (slices intercepting tumors) and BG slices was 10:90 % across all the three sets. Data augmentation (translation, rotation, scaling) was performed on the FG slices to get a balanced set between FG and BG slices.

We used an ImageNet-pretrained Faster RCNN model. Stochastic gradient descent was used to optimize a weighted sum of detection loss (cross-entropy) and regression loss for box offset (L1) for 80 epochs. For accepting a prediction, the intersection over union (IOU) threshold between ground truth and prediction was set to 0.5.

For FG slices, tight boxes were fit around the physician’s segmentation. Histograms were generated for the location and dimension of the boxes on FG slices. For BG slices, 3 different annotation methods were used:

M1 (whole slice as BG): The box center (x_c, y_c) was chosen as the center of the BG slice and dimension (w, h) was set nearly equal to slice dimension. Hence, for slice dimension normalized to 1, x_c=y_c=0.5 and w=h=0.9. This was our baseline. M2 (uniform distribution): The (x_c, y_c) were sampled from a uniform distribution that would assign values between (0.05, 0.95) and (w, h) were also sampled uniformly with well-defined upper bounds to keep the entire box inside the slice. M3 (tumor-box-like distribution): The (x_c, y_c), (w, h) were sampled from the histograms of the centers (Hist(X), Hist(Y)) and dimensions (Hist(W), Hist(H)) of the boxes on FG slice (Fig 1). The bin number was set to 50.

Detection performance was evaluated using detection accuracy (AC) and mean average precision (mAP) metrics for the 3 methods.

Results: We achieved AC=70% and mAP=0.66, AC=72% and mAP=0.67, and AC=74% and mAP=0.70 on the test set for M1, M2, and M3, respectively. M2 and M3 outperformed M1 as these were trained on various dimensions and locations of BG boxes in a randomized way; the model utilized a much more random sample to learn from in order to generalize better to the test set. M3 also outperformed M2. Since in M3, boxes were sampled from the distribution of boxes around tumors, the BG boxes were of similar sizes & in similar locations as the tumors and hence the model learned to identify the tumor well from the BG, reducing false negative & false positive detections.

Conclusions: In this work, we emphasize that a model’s ability to identify clinically non-useful classes (BG) can enhance the performance on detecting useful classes (tumors). We are working on employing various other BG annotation methods such as annotating non-tumorous high-uptake regions on the BG slices (to help reduce false positives), as potential steps towards improving automated deep-learning based detection model.

• Download figure
• Open in new tab
• Download powerpoint