Osteosarcoma (OSA) is the most common primary malignant bone tumor. Μultimodal treatment plan, including neoadjuvant chemotherapy, a wide-margin surgical resection of OSA and post-operative chemotherapy, is the gold standard for therapy of localized OSA. The prognosis of OSA is based on the necrosis rate following neoadjuvant chemotherapy. The extent of necrosis is estimated by pathological evaluation of the maximum cross-section of the tumor. The ratio of the necrotic area over the total tumor area can vary from <50% (grade I) to 100% (grade IV). A good prognosis is predicted if the necrosis rate is ≥90%. However, it has been shown that the necrosis rates are not reproducible and do not reflect individual cell response to neoadjuvant chemotherapy. Kawaguchi et al. in a recent manuscript published in the npj Precision Oncology journal, went on to investigate whether viable tumor cell density assessed using a deep-learning model (DLM) would reflect a more accurate prognosis of osteosarcoma. The benefits of using DLM is that the evaluations of viable tumor cells in resected specimens are very rapid and reproducible.
For the training of the DLM, authors used pathological images of 15 OSA patients. Following H&E staining and scanning of the specimens to generate whole slide images (WSIs), three pathologists annotated the nuclei of viable tumor cells by consensus, using an annotation software. With these annotated images the DLM was trained to detect the nuclei of viable tumor cells. To evaluate the DLM detection performance authors measured precision, recall, and F-measure, where precision refers to the proportion of viable tumor cells detected by the DLM that were also viable tumor cells in the annotated image; recall is the proportion of the cells of the annotated image that the DLM successfully detected as viable tumor cells; and the F-measure is the harmonic mean of precision and recall. The DLM’s detection performance was compared to that of one of the pathologists who generated the annotated image and repeated the annotation process again. Finally, the reproducibility of the detection performance of the DLM, was validated using specimens from OSA patients treated at an external facility.
Findings showed that the detection performance of the DLM did not vary significantly from the performance observed when the pathologist annotated the images twice (precision, recall, and F-measure were 0.82, 0.69, and 0.75, respectively), and when the external validation was conducted using OSA patient samples from an external facility, the results demonstrated evaluation metrics for each fold, with precision at 0.80 (SD 0.03), recall at 0.70 (SD 0.02), and F-measure at 0.75 (SD 0.01).
This work was the first to employ a DLM to calculate viable tumor cell density with a performance target set at a level comparable to that of humans. Authors concluded that DLM-evaluated viable tumor cell density is a more precise prognostic factor of OSA than the necrosis rate assessed by pathologists, and suggested that this approach has the potential to enhance the accuracy of prognostic stratification of osteosarcoma patients treated with neoadjuvant chemotherapy.