Opportunities and challenges in refining grading of neuroendocrine tumors using digital pathology

Grading of gastroenteropancreatic (GEP) well-differentiated neuroendocrine tumors (WD-NETS), based on the current WHO classification, depends on Ki67 proliferation index (PI) and mitotic index (MI) per 2 mm2. Accurate classification and grading are important for determining treatment strategies and prognostication of these tumors. The WHO recommends manual estimation as the gold standard. The criteria for performing manual PI are counting the number of positive cells in at least 500 tumor cells in the area of highest apparent proliferative activity (so-called hotspot region). For estimation of MI, the WHO recommends counting the number of mitotic figures (MF) in a 2 mm2 area in the most mitotically active area. Both are considered while assigning the final grade. The cutoffs for PI and MI are as follows: Grade 1 - PI<3% , MI<2 per 2mm2, Grade 2 - PI=3-20%, MI=2-20 per 2 mm2, Grade 3 - PI>20%, MI>20 per 2mm2.Estimation of PI and MI is cumbersome and subject to interobserver variation. The challenges in estimation of Ki67 PI start at choosing a block for staining, defining a positively staining cell, and determining a hotspot area, to name a few. Pathologists often choose a block showing more mitoses, atypia, or high-risk features; however, none of these may be present in a given tumor, making block selection arbitrary. Evaluation of Ki67 on multiple blocks is cumbersome and not cost effective. Choosing the right focus for MI estimation is subject to the limitations of the human eye as well as compromised by tissue processing/preservation artifacts and inflammatory cells.Our group has been working on using computational pathology to refine and simplify practices in grading of GEP WD-NETs. Use of image analysis makes it possible to study multiple tissue blocks to perform PI and MI for accurate grading. This is possible with both camera-captured images and WSI. Automated hotspot detection minimizes subjectivity in PI estimation. We have also compared multiple open-source platforms to perform PI on WSI and camera-captured images and found concordant grade assignment in most cases.As adoption of digital pathology gains traction, one would have to adapt to use of WSI for MI but there are no guidelines available for the same. We undertook a pilot study of MI estimation on WSI vs gold standard method of estimation on glass slides. Our methodology for MI estimation on glass slides includes low power panning of the slide to find hotspots. Once such an area is found, the first MF is counted and MF in next 10 consecutive 40X fields (2 mm2). On WSI, we tested two strategies - 1. Placing a grid overlay on the WSI, with each square measuring of 2mm2. This is followed by panning the slide on 10X to find any MF. Each MF is annotated and finally the maximum number of MF in any one square is recorded as MI. 2. Panning the slide on 10X, the first MF is annotated followed by drawing a square annotation measuring 2mm2 around the MF. Any additional MF in this square annotation are counted and the total number is recorded as the MI. 10 cases (4 small intestine and 6 pancreas) were selected for assessment by the 3 methods. We found excellent agreement between the glass and digital grid method (interclass correlation coefficient, ICC=0.97), digital square annotation (ICC=0.98) and between the two digital methods (ICC=0.99). There was no difference in final grade assigned by any of these methods. We found that MI estimation on WSI takes longer (average 8.3 min/slide for grid method, 7.3 min/slide for square annotation method vs 3.05 min/slide for glass slides). The reasons for this are multiple. Pathologists are not used to assessing MI on WSI. Inability to perform fine adjustment can limit definitive identification of MF. Scanning a WSI on 10X is cumbersome and time consuming. However, we did notice some advantages of using WSI. Counting in a standard area of 2 mm2 (either grid or annotation) is easier than counting MF in consecutive high-power fields. Furthermore, one can annotate the MF on WSI, enabling comparison of multiple hotspots and for future review. Machine learning algorithms can reduce the time spent in finding MF, aid accurate detection of hotspots and minimize subjectivity. Deployment of algorithms on sections from multiple tumor blocks can refine the grade assignment. However, developing these tools is also fraught with challenges - tumor cell detection, staining intensity, debris, fixation artifact and artifact exclusion are some of the common challenges. The balance between a sensitive and also specific algorithm is often hard to achieve.The aim of this presentation is to share our experiences and lessons we have learned. Much work remains in developing a standardized, consistent, and accurate approach to grading GEP WD-NETs, but we believe that using image analysis will be critical in refining grading of these tumors and using open-source technology will be very helpful in ensuring wider applicability and uniformity in grading.

Learning Objectives

1. Enumerate the available open-source options for performing grading of GEP-NETs.
2. Understand the advantages of using image analysis for grading of GEP-NETS and choose the best option based on available resources.
3. Understand the challenges and limitations of using image analysis for grading of GEP-NETS.