Universal Lymph Node Detection in Multiparametric MRI with Selective Augmentation

TL;DR

The paper tackles the challenge of universally detecting lymph nodes in multiparametric MRI across heterogeneous scanners and protocols to support subsequent sizing and cancer staging. It introduces a VFNet-based detector augmented with Intra-Label LISA selective augmentation to improve robustness when fusing T2-weighted fat-suppressed and DWI inputs in 2.5D configurations. The best results come from a 2T2FS+1DWI input with ILL, achieving a mean average precision of about 56% and recall of 82.4% at 4 FP/vol (89.2% at 8 FP/vol), outperforming several prior methods and markedly increasing recall relative to T2FS-only approaches. The work demonstrates a fast, practical LN detection pipeline (≈2.9 seconds per volume) that could streamline radiologists’ workflow by enabling reliable LN localization for measurement while accommodating real-world data diversity; future work includes adding more mpMRI sequences like ADC and expanding LN mining to further improve performance.

Abstract

Robust localization of lymph nodes (LNs) in multiparametric MRI (mpMRI) is critical for the assessment of lymphadenopathy. Radiologists routinely measure the size of LN to distinguish benign from malignant nodes, which would require subsequent cancer staging. Sizing is a cumbersome task compounded by the diverse appearances of LNs in mpMRI, which renders their measurement difficult. Furthermore, smaller and potentially metastatic LNs could be missed during a busy clinical day. To alleviate these imaging and workflow problems, we propose a pipeline to universally detect both benign and metastatic nodes in the body for their ensuing measurement. The recently proposed VFNet neural network was employed to identify LN in T2 fat suppressed and diffusion weighted imaging (DWI) sequences acquired by various scanners with a variety of exam protocols. We also use a selective augmentation technique known as Intra-Label LISA (ILL) to diversify the input data samples the model sees during training, such that it improves its robustness during the evaluation phase. We achieved a sensitivity of $\sim$83\% with ILL vs. $\sim$80\% without ILL at 4 FP/vol. Compared with current LN detection approaches evaluated on mpMRI, we show a sensitivity improvement of $\sim$9\% at 4 FP/vol.

Figures (2)

• Figure 1: LN detection results of a VFNet model are shown in (a)-(f). Green boxes: ground truth, yellow: true positives, and red: false positives. (a) shows a T2FS slice and (b) shows a DWI slice from a mpMRI study. (c) and (d) show results from a 2 T2FS + 1 DWI slice combination; (c) shows results without Intra-Label LISA (ILL) on a T2FS slice, while (d) shows results with ILL on an interpolated slice. Notice the prominent granularity of the spleen visible in (d) due to interpolation of two domains (T2FS and DWI). (e) and (f) display results from a 1 T2FS + 2 DWI slice combination; (e) shows results without ILL on a DWI slice, while (f) shows results with ILL on an interpolated slice. Note that FP were detected in all images except for (d). (g) and (h) show the results of a comparative Faster RCNN baseline Zhao2020_mri on a 2 T2FS + 1 DWI slice combination; (g) shows results without ILL on a T2FS slice, while (h) shows results with ILL on an interpolated slice. Note that a LN was missed in (g) and (h) that was captured in (d). VFNet trained with ILL showed higher recall with fewer FP.
• Figure 2: Output of the VFNet model on slices from different mpMRI studies. The model was trained with the 2 T2FS and 1 DWI slice combination. The top and bottom rows show outputs of VFNet in $E_{21}$ trained without and with ILL respectively. The top row shows only T2FS slices, while the bottom row shows interpolated slices. Green boxes: ground truth, yellow: true positives, and red: false positives. LN of different sizes (SAD $\geq$ 3mm) that were missed in (a), (c) and (e) were captured in (b), (d) and (f) respectively. (b) and (h) also saw a reduction in the number of FP in contrast to (a) and (g).