In this study, combining RI imaging data with color Pap staining imaging data improved the accuracy of MLA in cancer diagnosis using thyroid FNAB specimens. The results of MLA classification using color Pap staining images were highly dependent on the size of the nucleus, but the results of MLA classification using RI images were less dependent on the size of the nucleus, indicating that the I was affected by information. The final algorithm, which used data from both types of images together, distinguished thyroid cell clusters from benign thyroid nodules and PTCs with 100% accuracy.
MLA employs a convolutional neural network (CNN) architecture, which is effective for image analysis, and has shown excellent diagnostic performance using images of thyroid FNAB specimens.7, 8, 12, 13. Guan et al.13 We studied a CNN-based MLA for classifying hematoxylin-eosin-stained FNAB specimens of benign thyroid nodules and PTCs (TBSRTC II, V, VI). A total of 887 fragmented color images were used in this study, cropped from his 279 images taken using a digital camera attached to a brightfield microscope. The trained algorithm showed 97.7% accuracy in differentiating 128 test images of benign and malignant nodules. Range et al.8 We used MLA to classify Papanicolaou-stained FNAB specimens of extensive thyroid nodules (TBSRTC II-VI). They used 916 color images acquired using the entire slide scanner. Her trained MLA distinguished between malignant and benign nodules with high accuracy (90.8%) comparable to pathologists. Similarly, his CNN-based MLA showed high-accuracy patch-level classification (97.3%) and cluster-level classification (99.0%) using only color Papanicolaou-stained images, performing well in our study. showed.
However, given that the purpose of FNAB is to decide whether to operate a thyroid nodule, not only does it show high overall accuracy, but it also classifies an overt malignant tumor as should minimize significant misclassification, such as classifying as benign. Malignant tumor. In Guan’s study, the MLA incorrectly classified some cases as malignant that pathologists classified as clearly benign. Similarly, in Dr. Renji’s study, the MLA misclassified some apparently benign nodules as malignant, and misclassified malignant nodules requiring surgery as benign.8. These issues are problematic because they can lead to incorrect treatment planning for patients who would otherwise be well treated under the current standard of care. We studied nodules (TBSRTC II, V, and VI) with relatively different benign or malignant features. Our finding that RI data improved the accuracy of MLA in these nodules has important clinical implications as it indicates a potential reduction in the aforementioned severe misclassification.
Guan et al.13 suggested that significant misclassification of MLA in thyroid FNAB specimens may be related to nuclear size. In their study, cells in false-positive cases exhibited large nuclei with high average pixel color information, similar to malignant cells, but pathologists determined that these cells typically had a benign morphology. Did. The authors interpreted that MLA classification was based on nuclear size and staining, not on shape. Moreover, in our results, since nuclear size is the main feature required for classification, MLA based on color images can accurately classify benign thyroid cells with large nuclei or malignant thyroid cells with small nuclei. has been shown to have limitations. However, MLA classification based on RI images was less affected by nuclear size. This suggests that RI imaging can compensate for the limitations of MLA, using color images of FNAB specimens whose nuclear size is not typical of benign or malignant cells.
Further analysis results to explain the model suggest that MLA based on RI images uses nuclear structure and shape for classification. In color images, the algorithm was activated mainly for large nuclei, whereas in RI images, the algorithm was activated not only for large nuclei but also for well-structured nuclei. The reliability of the MLA classification results was proportional to the detail of information around the nuclear envelope when based on RI images, but not when based on color images. Detailed nuclear structures such as nuclear envelope irregularities and micronuclei are important indicators for thyroid cancer diagnosis.26. Therefore, incorporating such information can improve the accuracy of MLA classification.
Another potential strength of RI images is their ability to integrate information in a wide vertical space. In thyroid cytology specimens, the cells are scattered over a wide vertical space (i.e., multiple Z-planes) rather than on a plane. Monolayer (z-plane) 2D images cannot cope with this vertical spread, and information from out-of-focus cells can be lost or distorted. In contrast, in RI images acquired by ODT, cells located in different Z-planes are in focus simultaneously. In our study, color image-based MLA showed false positive results for some out-of-focus patches, whereas RI image-based MLA showed true negative results for the same image patches. (data not shown). However, the out-of-focus region is only a part of the color image, and compared to using a single Z-plane image in previous studies, the accuracy of MLA is still lower even when using multiple Z-plane images. Did not improve.8. Therefore, it is unclear whether the aforementioned factors significantly affect MLA accuracy.
This study has certain limitations. Despite the large number of sample measurements, this study was conducted at a single center and therefore could not cover all sample conditions that may exist in a real clinical setting. ODT provides optimal RI imaging for unmanipulated live cells27, but we acquired RI images from colored-stained cells. Staining acted as extraneous noise or artifacts within the RI images and reduced the accuracy of MLA. Further studies are needed to determine the effect of staining on results. Finally, up to 30% of FNABs may exhibit ‘indeterminate’ cytopathology (TBSRTC III and IV). Because this study targeted specimens characteristic of benign or malignant thyroid nodules (TBSRTC II, V, and VI), we could not apply the currently trained algorithm directly to TBSRTC III and IV specimens in the absence of relevant training. Not applicable.
To explore the complementary nature of RI and color images, 2D MIP images were generated by projecting 3D RI images along the Z axis, thereby eliminating the effect of dimensionality. Previous studies in the field of cell classification have demonstrated improved performance using 3D RI images compared to 2D images.28,29. In our study, his 3D images were not incorporated due to specific research purposes, but future studies plan to extend the study by incorporating 3D RI images and other his 3D imaging modalities.
In this study, we demonstrated the efficacy of multiplexing RI with standard brightfield imaging using a single ODT platform for MLA-based classification of benign and malignant thyroid FNAB. Multiplexed ODT has shown promise in developing a more precise classification of thyroid FNAB while mitigating the inherent uncertainties and errors observed in current diagnostic standards. Therefore, ODT-based his MLA could potentially contribute to more cost-effective and rapid point-of-care management of thyroid malignancies.
