Author: Deepbio

Authors:
JaeHeon Lee, Tae-Yeong Kwak, Joonyoung Cho, Sun Woo Kim, Hyeyoon Chang, Hong Koo Ha

Organizations
Deep Bio Inc., Seoul, South Korea, Pusan National University, Pusan, South Korea

Background
After radical prostatectomy (RP), a steep increase in PSA level is an early sign of the disease progression in prostate cancer, which is known as biochemical recurrence (BCR). The risk of BCR can be evaluated based on a combination of clinicopathological factors, and the patient’s Gleason score plays a significant role. However, this scoring system may have lower consistency due to interobserver reproducibility in classifying Gleason Patterns (GP) as well as in quantifying the amount of each GP. We developed AI-based nomograms, with the aim of investigating their prognostic efficacy.

Methods
In this study, digitized whole-slide images (WSIs) of H&E-stained prostatectomy specimens and clinical follow-up information were obtained from two sources: Pusan National University Hospital (PNUH, n = 967, event = 342) from 2010 to 2021 with the median follow-up being 3.7 years, and The Cancer Genome Atlas (TCGA, n = 352, event = 79) from 2000 to 2013 with the median follow-up 2.6 years. We used the DeepDx Prostate – RP, an AI-based prostate cancer Gleason grading model, to compute a pixel-wise probability map of each GP in WSIs. Then, a weighted sum of the probabilities and GPs was calculated for each pixel. The proposed slide-level score (AI score) was then determined by averaging them across all pixels in WSI, resulting in a value ranging from 3 to 5. Also, patients were divided into five groups using AI score thresholds (3.1, 3.5, 3.9, and 4.1), and five binary variables (AI score group 1-5) were generated, where a value of 1 indicates a patient’s categorization into a group. To evaluate predictability we created new nomograms incorporating AI scores based on existing nomograms: MSKCC and CAPRA-S. Unlike the original nomograms, the proposed nomograms do not include the Grade Group (GG) made by pathologists. Then, we fitted a Cox regression model on one of two datasets using the original and newly formed nomograms, and validated on the other dataset reporting the concordance index (c-index). Confidence intervals for the c-index were generated via the non-parametric bootstrap resampling with 9,999 samples.

Results
The table shows the prognostic performance of sets of clinicopathological factors and demonstrates that the proposed AI score improved the predictive power. The nomograms with AI score group 1-5 outperformed the original nomograms, as shown in (d).

Conclusions
In conclusion, we developed the AI-based nomograms which can improve the accuracy of predicting the biochemical recurrence in prostate cancer compared to the existing nomograms.

Authors:
JaeHeon Lee, Tae-Yeong Kwak, Joonyoung Cho, Sun Woo Kim, Hyeyoon Chang

Organizations
Deep Bio Inc., Seoul, South Korea

Background
Prostate cancer affects millions of men globally and biochemical recurrence (BCR) is an important indicator of its progression. With the increasing use of computational techniques, there has been a growing interest in utilizing whole slide images (WSIs) of H&E stained prostate tissue to evaluate the patients’ clinical risks. Our study aimed to develop a novel method for predicting BCR by performing clustering on features extracted from deep learning models, thereby capturing morphological characteristics.

Methods
The WSIs and clinical information were collected from The Cancer Genome Atlas (TCGA, from 2000 to 2013) dataset, for 321 patients (43 patients with BCR) who were not treated with any adjuvant therapies, with 2.5 years of median follow-up time. 356 collected WSIs were split into 1024 x 1024 patches at 5x magnification for the deep learning-based analysis. We used an ImageNet-pretrained deep learning-based clustering model (SwAV) to extract features from each patch and applied additional K-means clustering to assign the patches into one of 16 clusters. We tested both the cosine similarity and Euclidean distance as the clustering measure. After excluding the dissimilar patches under specific thresholds, the proportion of each cluster was calculated per slide to describe the distribution of morphological patterns as a 16-dimensional vector. To evaluate the predictability, we trained Cox models with the proportion vectors and assessed their performance via the concordance index (c-index). We repeated a stratified 5-fold cross-validation process 100 times to achieve robustness. As a reference, we conducted the same experiment with a BCR prediction nomogram published by MSKCC. Moreover, with the median proportion of each cluster, we separated two patient groups, drawed Kaplan-Meier curves, performed the log-rank test for significance, and we obtained the hazard ratio (HR) of each cluster from the trained Cox model of each fold.

Results
We explored several thresholds to find ones yielding the best performance. When using the similarity, the best c-index was 0.704 with the threshold value of 0.25. With the distance value of 0.60, the c-index was 0.701. When using the MSKCC nomogram that incorporates clinical features, the average c-index was 0.717. The table contains the results of statistical analysis for selected clusterings that were significant in the log-rank test. Particularly in cluster B, the patches showed high-risk patterns including perineural invasion, high grade cancer, and necrosis in agreement with the high HR.

Conclusions
Our results suggest that the histomorphological pattern profiles depicted by the deep learning model have the potential for effective risk stratification in prostate cancer.

Authors:
Tae-Yeong Kwak, JaeHeon Lee, Joonyoung Cho, Sun Woo Kim, Hyeyoon Chang

Organizations
Deep Bio Inc., Seoul, South Korea

Background

Authors:
Tae-Yeong Kwak, JaeHeon Lee, Joonyoung Cho, Sun Woo Kim, Hyeyoon Chang, Hong Koo Ha

Organizations
Deep Bio Inc., Seoul, South Korea, Pusan National University, Pusan, South Korea

Background

The histologic grade (Gleason score) as well as the tumor ratio of the resected specimen plays a significant role in assessing the clinical risk of prostate cancer patients who underwent radical prostatectomy. Gleason scoring is known to be prone to inter- and intra-observer discordance, and eyeballing measurement of tumor ratio is possibly coarse and inaccurate. Several deep learning-based tissue image analysis algorithms that perform prostate cancer diagnosis have been developed and are about to be applied clinically. We evaluated the utility of one of those algorithms by comparing its output with the pathology reports.

Methods
A total of 29681 H&E-stained tissue slides were collected for 1001 radical prostatectomy cases during 2010-2021 at Pusan National University Hospital and scanned at 40x magnification into whole-slide images. On each slide, we utilized a tissue detection algorithm to identify tissue regions and measure the area. A deep learning-based algorithm was used to identify prostate cancer lesions, measure the area, and determine the grades. The area measurement was converted into the volume figure assuming that the tissue slice thickness was 5mm. The slide-wise algorithm outputs were then aggregated for each case, resulting in the specimen volume, tumor ratio, and Gleason score. In the evaluation, the algorithm-based Gleason scores were compared with the ones in the hospital pathology reports on the ISUP grade group basis. The correlation analysis was performed for the tumor ratio. The specimen density values were calculated from the volume figures and the hospital weight measurement and utilized to exclude the cases with extreme density values.

Results
The min, max, and average number of slides per case were 1, 67, and 29.7. For two cases with a single slide collected, both the algorithm and the expert pathologist found no cancer. The pathologist also confirmed that 7 had residual tumors among 15 cases where there was no Gleason score in the pathology reports while the algorithm found cancer. For the remaining 984 cases, the algorithm provided the grade groups equal to, higher than, and lower than the pathology reports for 482, 384, and 118 cases respectively, showing moderate agreement. The min, max, and mean values of the specimen volume (mL) were 0.01, 75.3, and 32.7, while the corresponding values of the specimen density (g/mL) were 0.11, 3001.79, and 4.38. The density values at the lower and upper 2.5% were 0.42 and 1.40 respectively, and the mean of the values between them (inner 95%) was 0.92. The degree of tumor ratio correlation between the algorithm and the pathology reports was high with the coefficient 0.813 (95% CI: 0.791-0.833), which went up to 0.826 (95% CI: 0.805-0.845) for the inner 95% density cases.

Conclusions
Our findings support the usefulness of the algorithm-based analysis of prostatectomy specimens, which can benefit clinicians in the hospital.