Alberta Precision Laboratories has a robust digital pathology slide set, used primarily for teaching. This slide set contains images of previous cases (hematoxylin and eosin–stained histology slides) that have been scanned using an Aperio GT450. The bulk of our case images came from this dataset. In addition, to increase variability within our dataset, both for hematoxylin and eosin stain quality, as well as scanner type, we also used publicly accessible WSIs from Leeds University, University of Michigan, and the Cancer Imaging Archive [25-27]. Cases (n=494) were randomly divided into 3 allocations: training (75%), validation (10%), and test (15%, Table 1).

In Table 1, cases and their associated images were split into 75% training, 10% validation, and 15% test. Slightly increased test cases were used (that the traditional 80/10/10 split) to better evaluate model performance, and to compensate for a relatively small training set.

We focused on 3 common entities seen in colon polyps: hyperplastic polyps, tubular adenomas, and normal colon. All cases were opened in Aperio Imagescope, and manual image/patch extraction was completed at 4× and 10× objective power (40× and 100× magnification, respectively), focusing on the most representative tumor/diagnostic areas. A total of one 4× extraction/tile and two 10× extraction/tiles were used per case, for a total of 1482 images (Table 1). Tiles were chosen by a pathologist to ensure that the most representative areas of the slide were used. The extraction was carried out using Aperio-Imagescope’s built-in image extraction tool, using the embedded International Color Consortium (ICC) profile and exported in .jpeg format. The ICC profile is a method of color normalization used by Aperio (and other digital pathology vendors) to ensure the image generated from the slide matches as closely as possible to the real-world color profile of the slide. Both the scanner and image profile used in this study are fully validated and accredited for clinical use.

Images in the dataset were rescreened by a blinded pathologist to ensure that the appropriate diagnostic label had been applied and to ensure that no images from test cases were present in either the training or validation data. A total of 3 users (2 pathologists and 1 pathology resident) were involved in the tile and review process to prevent bias in the tile selection. Instructions for tile selection were to select the most appropriate tile for the diagnosis. Only tiles with perfect consensus (3 of 3 agreed) labels were chosen to be used.

A single-label image classification model with 3 labels (“hyperplastic polyp,” “tubular adenoma,” and “normal”) was then developed using Vertex AI and an autoML algorithm. Other model parameters were set to default to demonstrate a general yet easy to use model for nontechnical experts (model details are as follows: training method: “AutoML,” Objective: “Image classification (single-label),” Data-split: “Manual,” Budget: 48 node hours, Actual: 43.04 node hours, Training time: 5 h 33 min).

The AI model was tested on the “test” allocation of cases that were not part of the training or validation datasets. XRAI overlay [28] was used to view explainability and ensure that the algorithm was identifying the pertinent areas of the image (Figure 1). Model output was evaluated on a per-image basis, as opposed to a per-case basis. Results from the test data were used to evaluate precision and recall. VertexAI incorporates model evaluation and automatically generates both the precision-recall curve as well as precision-recall by threshold. An overall confidence threshold of 0.50 was chosen for the evaluation of our model.

Ethics approval and a waiver of consent were obtained for using deidentified case images from our institutional database. This study was approved by the Health Research Ethics Board of Alberta (HREBA.CC-23-0347).

The integration of AI in the field of digital pathology is not merely a technological advancement; it is becoming a necessity driven by the contemporary challenges that the medical community faces. Here we have demonstrated that it is relatively easy to train your own AI algorithms on your own data, as a pathologist. As ML continues to revolutionize diagnostic pathology, our study demonstrates how simple off-the-shelf AI tools can readily be used to develop effective models for improving diagnostic efficiency. The recall and precision of the developed AI model in detecting colon polyps were remarkable, approaching 100%. This exceptional performance underscores the potential of ML to transform diagnostic pathology, especially given the rapid advancements in technology, availability, and cost-effectiveness of MLaaS platforms such as Google’s VertexAI.

Previous AI applications in digital pathology, specifically in colorectal cancer screening, have shown great potential in increasing the efficiency and accuracy of diagnosis. For instance, Korbar et al [29] demonstrated that deep learning models could classify colorectal histology slides with high accuracy, bridging the gap between manual microscopic evaluations and automated assessments. Our study builds upon these foundations, offering further evidence for the efficiency and effectiveness of AI in this domain. One could say we have crafted a model that emulates the performance of an early-stage pathology resident, as the 3 entities used in this model are relatively simple; the difference is that this model took days to train, versus years for the average pathology resident. Nevertheless, there are certain limitations to consider.

First, our study had a limited sample size and did not include a complete range of diagnostic entities. With only tubular adenoma, hyperplastic polyp, and normal colon as labels, this model does not account for other critical entities like serrated adenomas, high-grade dysplasia, or carcinoma, an important histologic feature of high-risk polyps [30]; this would be a good area for future projects. AI models, particularly deep learning models, generally require large training datasets in order to generalize well in real-world scenarios [18,31]. A more extensive dataset might allow for more refined models that capture nuances and rare presentations of colon polyps, such as high-grade dysplasia. This limitation is significant, as evidenced by the emphasis on broad scope in successfully used deep neural networks for detecting colorectal cancer on WSIs [32].

Second, while our model shows promise, integrating it into real-world workflows would necessitate the inclusion of a more diverse range of clinical criteria beyond just the image data. We did not include certain clinical characteristics that would be important, like location (ie, ascending colon), which can affect diagnosis, especially for serrated lesions [23].

Third, the need to manually annotate our dataset means that we are effectively working with a “best-case scenario” data input. Real-world scenarios might present slides with artifacts, suboptimal staining, or overlapping tissues that could challenge the model’s predictions [33]. The success of AI models in clinical settings largely depends on the quality and diversity of the input data. As we only used tiles with consensus between our 3 subject matter experts (2 pathologists and 1 pathology resident), this meant a small number of imperfect tiles were excluded. These “imperfect” samples, of course, exist in the real world, and these must be interpreted as well.

Fourth, our model misclassified a number of “normal” as “hyperplastic polyp,” which appeared to be on normal colon where glands were not perfectly round. Likely inclusion of more “normal” images with different-shaped lumens would help tune the model for proper identification.

However, despite these limitations, our study exemplifies the promise of using ML as a service in histopathology. The near-perfect performance of our model in a controlled environment suggests that with adequate training data and a broader range of diagnostic entities, such platforms could soon play a significant role in augmenting the capabilities of pathologists.

Our project has shown that widely available “AutoML” algorithms, such as Google’s Vertex AI, can be applicable to the medical field, specifically in digital pathology. Our training model was robust with a sensitivity approaching 100% and a specificity of 98%. This aligns with other authors’ findings of success with AutoML algorithms in the medical field [31]. Such AI tools will help to improve the efficiency of front-line pathologists, allowing them to keep up with increasing demand and so helping to mitigate workforce constraints. For example, our model could be used to auto-generate a report, saving time for the pathologist in dictating/typing.

Overall, we show that cloud-based ML platforms can produce accurate models that are specific to the field of pathology and WSIs. Furthermore, they are easy to use, with only a relatively basic level of ML knowledge required. As the medical community strives for implementation of precision medicine that should lead to improved patient outcomes [34], integration of ML tools into the realm of histopathology may soon become indispensable. Future studies with larger datasets, diverse diagnostic entities, and more real-world scenarios are warranted to further elucidate the potential of AI in this domain.