Commonly (mis-)used terminology

Training and overfitting

Different mathematical machine learning models exist to describe the relation between an input (eg, image) and a desired output (eg, a label). Support vector machines, random forests or neural networks are popular model choices. Although different machine learning models are constructed in different ways, all of them use data for building the model. Building such a model based on data is referred to as training, since the model learns from given examples. During training, driven by mathematical optimisation methods, a model will gradually improve in capturing the input-output relation of the given examples in the training data. For example, given a set of images of polyps and the associated class of these polyps, it will learn to predict the class, based on the image. Once a model is trained, it should also work for new data, which was not used for training the model. This property is called generalisation. If the model only works for the training data, but does not work well for new data, this is called overfitting, as the model is too tightly fitted to the training data and does not generalise towards new data (figure 1). Especially for smaller sets and more complex models, overfitting poses a serious problem.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Graphical display of overfitting of training data. In this figure, the leftmost panel displays data points of two classes, in which the class is indicated by the colour. The centre panel shows the same data including the prediction of a model trained on that data as the background colour. Overfitting is clearly visible as the model isolates points of the red class, rather than capturing the class as a whole. The rightmost panel shows the prediction of a different model as background colour. Although this model makes mistakes (red points can be seen on a blue background and vice versa), this model demonstrates better generalisation, as it captures the class distributions rather than individual points.

Training, validation and test sets

In building a machine learning model in endoscopy, the available data are generally split into three distinct sets: (1) training set, (2) validation set and (3) test set (figure 2). The training set is used to build the model (eg, to train a model that predicts a label based on image features). To check that the model is not overfitted to the training data, the validation set is used. Using this separate set, one can verify that the model also works for unseen data. In addition, almost all machine learning models are in part regulated by so-called hyperparameters, for which optimal choices can also be selected based on the performance on the validation set. Finally, a third part is used to evaluate the model that is trained with the optimal hyperparameters. This part of the data is referred to as the test set. It is important that the hyperparameters are not selected based on the model’s performance on the test set, as this would bias the model towards the unseen data and thus create a type of overfitting. This effect is also known as data leakage, since data that should be used only to test performance are also used to optimise the model.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Visualisation of training, validation and test set and overfitting, and their appropriate use. The training dataset is used to train the model, followed by validation. In case of unsatisfactory performance, the model is changed, retrained and again validated. In case of satisfactory performance, the model is then tested on a separate test set to evaluate model performance.

Cross-validation

As the availability of annotated endoscopic data is often limited for training and validating CAD systems, a specific validation procedure can be used to optimally exploit the available data, which is called cross-validation. In this validation approach, data are split into a number of equal parts, most often four, after which one part is used as a validation set, while the remaining (in this case three) parts are used for training. This is repeated a total of four times, each time with a different validation set, after which the scores of the four experiments are pooled into a point estimate, yielding a result that is more robust against data variation. Figure 3 graphically displays a dataset with fourfold cross-validation. When using cross-validation, it is important that the partitioning of the data is performed on a patient basis: data from a single patient can only occur in one of the four parts. When this approach is not carefully followed, cross-validation will likely provide a biassed overestimate of the algorithm performance. For example, when researchers present cross-validation results where different polyps from the same patient are included in multiple ‘folds’ (ie, both in training and validation parts), this will likely lead to an overestimation of the model performance. A good way to deal with an unbalanced dataset, for example, by including many polyps from a single patient versus only a small number of polyps from other patients, is to stratify the dataset in size and class balance. Note that, after optimising the hyperparameters (figure 2) using cross-validation, still, an independent test set is necessary to estimate the algorithm performance.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3

Graphical display of fourfold cross-validation.

Selection bias, overfitting and representativeness of data

It is important to realise that machine learning algorithms will pick up on any discriminative feature that allows them to better separate the data, irrespective of whether this discriminative feature is logical, clinically relevant or clearly a result of bias.13 A great example is a study in which an algorithm was developed to recognise neoplasia in histology slides. The algorithm appeared to be highly effective at first sight, yet in fact was found to recognise ink marks placed by pathologists to indicate abnormal areas rather than actual neoplastic morphology.

Most machine learning studies in endoscopy have used retrospectively collected datasets, often collected in previous studies with strict selection criteria.17 These datasets are generally retrieved from single-centre databases in expert centres, containing many similar samples that do not cover the natural variability of imaged objects, endoscopists and imaging devices in daily practice.18 These studies are therefore prone to selection bias (ie, the non-random selection of data that does not fully represent the study population and thereby limits the external validity of results) and overfitting, a common phenomenon in machine learning.13 Overfitting leads to overestimated results that lack representativeness in a general endoscopic setting. The risk of overfitting is further increased when only static, endoscopic images are used, since endoscopists tend to only capture and store high-quality images. Videos, encompassing 30 video frames per second, often contain a larger variety in image quality (eg, by the presence of blur, stool or partially obscured lesions).8 10 Thereby video analysis is likely to add to the robustness of CAD systems, lowering the risk of overfitting, compared with an image-based approach.

A general rule to minimise the risk of overfitting is to use datasets that are both large and heterogeneous. The use of multiple heterogeneous datasets is preferred, since heterogeneous training leads to robustness and is one of the best measures to prevent overfitting. Such datasets should ideally be independent, meaning that they at least exhibit no overlap in terms of patients, but preferably also vary on other aspects such as the operator, medical centre or acquisition protocol. Prospectively collected datasets are less prone to selection bias and are often more similar to a general endoscopic setting. That said, the collection of large prospectively collected datasets is cumbersome. Nevertheless, prospectively collected datasets should be preferred over retrospectively collected datasets in test sets.

To enable identification of bias and overfitting in CAD studies, it is important that all datasets used in the process of training, validation and testing are well defined. A complete description of datasets encompasses at least: data source (eg, random search in endoscopic database, existing study cohort, prospective protocol); inclusion and exclusion criteria of imagery; number of collected and excluded images and number of patients. Finally, some basic technical information of imagery and any preprocessing techniques should be stated (eg, resolution, file type).19 Moreover, it is important to consider these technical aspects when planning to construct such a dataset, as they can negatively impact the results in a real-time setting and even lead to overfitting. For example, if a lossy compression standard such as JPEG is being used in a multicentre study, it is likely that slightly different compression settings are employed for each centre. An algorithm could exploit this bias to learn that images with a certain compression type are more likely to be of a certain class. This is an especially notorious type of bias, as modern compression standards are optimised to produce artefacts that cannot be perceived by the human visual system. Table 3 shows a complete overview of basic data requirements.

Gold standard development

Next to the availability of multiple independent datasets, which should preferably be large and heterogeneous, the reliability of the label of interest is of crucial importance. We can only build a reliable model if relevant predictive features are related to an outcome which is ‘true’. This implies that for all the datasets we are using to build the model (for learning, validation and testing), the label of interest should be unequivocally established. This is called the gold standard (or ground truth) input for the algorithm. For CADx algorithms, the gold standard input is generally the histology of the corresponding image. Such a classification model then allows for image-based prediction of the histology of a new image, for example, normal versus neoplastic, or adenomatous versus hyperplastic polyp.10 20 21 It should however be noted that histology might not always be the preferred gold standard for CADx systems, for example, due to sampling error and interobserver variability between pathologists.22 For example, in certain circumstances, optical biopsies by experts might be preferred over pathology assessment.

Providing a suitable gold standard for localisation algorithms is more challenging, since these have to classify imagery, and generally have to localise lesions within imagery. This latter constraint holds for most CADe algorithms, therefore they are generally trained with manual annotations. However, for localisation of a lesion within an image, it is nearly impossible to obtain a pixel-precise annotation based on the pathology findings. Therefore, in addition to a histopathological gold standard label of the image, delineations by multiple endoscopists offer a reasonable estimate of the location of the lesion within the image. This is a laborious process, generally performed by expert endoscopists.

For this reason, often a single expert annotation is used for training and/or testing.18 23 24 It is important to realise that such gold standard annotations are only an approximation of the actual gold standard, that is, the demarcation line of underlying pathology. Since experts tend to disagree on pixel-precise annotations, causing interobserver variability, single expert delineations are subjective and less accurate. By including multiple expert delineations for each image, it is possible to differentiate between ambiguous image areas and areas where there is expert consensus. Especially for targeted biopsies, the expert consensus provides a valuable estimate of the optimal biopsy location. Figure 4 displays an exemplary image of subtle Barrett’s neoplasia for which three experts were asked to indicate the borders of the lesion. Although all three experts agreed on a central part of the lesion (‘the sweet spot’), they disagreed on other parts. The ‘sweet spot’ logically has a higher likelihood to contain neoplasia, and therefore contains more information and thus serves as a better gold standard than the individual expert delineations.25 It should be underlined that the localising function of such CAD algorithms only serve to improve the primary detection of neoplastic lesions in overview, allowing targeted biopsies by non-expert endoscopists who otherwise would have missed the lesion. An exact histologically correlated delineation is not required since the actual resection of the lesion will be done using a different endoscopic approach (eg, optical chromoendoscopy with optical magnification), often in another endoscopic session and by a more experienced endoscopist.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4

Exemplary case of subtle Barrett’s neoplasia, delineated by three experts (yellow, blue and green). Parts of the lesion (‘the sweet spot’) are recognised by all experts (black), yet other parts are only recognised by one or two experts. Reprinted from Bergman J, de Groof AJ, Pech O, et al. An interactive web-based educational tool improves detection and delineation of Barrett's esophagus-related neoplasia. Gastroenterology 2019;156:1299-1308, with permission from Elsevier.

In addition to assessing the credibility of the gold standard predictions, exploiting interobserver variability is a growing topic of interest within the machine learning community and could help in making more robust models.26

The use of multiple expert annotations per case is therefore preferred in establishing a gold standard for CADe algorithms. To minimise interobserver variability of such gold standard annotations, the annotation instructions for the experts should be determined in advance to avoid a scenario whereby one expert annotates only the gross abnormal part of the lesion yet the other expert attempts to precisely delineate the more subtle periphery of the lesion. In publications, the quality of the gold standard annotations should be clearly described (number of experts, their qualifications and their pre-delineation instructions).

Delineation of videos, containing thousands of video frames, is often performed by non-experts. It is imperative that these delineations are confirmed by experts to ensure accurate and correct gold standard establishment.

Data ownership

The use of clinical patient data for development of (commercial) CAD systems is currently under debate in the world of AI, involving both ethical and regulatory issues. This debate exceeds the field of machine learning in endoscopy but is of vital importance, since all research groups that are using clinical patient data without specific consent will be confronted with local and international legislation when they are seeking commercialisation of their product. There currently is an unmet need for uniform interpretation of legislation on this topic.27 Pending a commonly accepted legal interpretation of this issue, authors should provide detailed information on ethics approval in their scientific papers.

The chosen performance metrics and their motivation

There are different ways to express the performance of a model. For the different machine learning tasks, several well-defined metrics have been described: sensitivity, specificity and area under the curve for two-class classifications, confusion matrices and mean average precision for multiclass classifications and intersection over union (IoU) or the DICE coefficient for segmentation (ie, delineation).29 However, sometimes these metrics do not fully reflect the desired algorithm outcome. For example, when considerable interobserver variability is present around the borders of gold standard annotations, these areas may be considered of lesser importance than the area for which there is consensus, as described in the previous paragraph. In that case, it is not obvious to apply IoU or DICE, as there are multiple segmentations that are correct, and these metrics can only deal with a single gold standard segmentation per image. In that case, one can use a slightly adapted version of these metrics25 30 or aim to explicitly model the interobserver variability and incorporate this into the evaluation metric.31 When deviating from the set of default metrics, it is important to provide a clear rationale, to avoid metrics being presented just because they yielded the most favourable results, which would then lead to selection bias and favour overfitting.

Approach for validation and testing

Another crucial parameter in the experimental design of machine learning algorithms is the chosen approach for validation and testing. As mentioned, machine learning models typically contain several hyperparameters, which can be used to optimise its performance. To test the trained model’s performance, generally a separate test set is used, containing data that not have been used for training the model. While running the trained model on this test set, it may be tempting to adjust the hyperparameters of the existing model based on its performance on the test set. This leads to a form of overfitting generally known as ‘data leakage’ (as information contained in the test data is actually used (ie, ‘leaks’) to train the model instead of only to test its performance). Data leakage thereby often leads to a misrepresentation and overestimation of the model’s actual performance.

To avoid data leakage, usually a validation set is used as an intermediate step in the experimental design. This validation step is then used for optimising the hyperparameters of the model, and thereby preventing data leakage from the test set. In this ‘training set-validation set-test set sequence’, the training set is used to optimise the model’s performance in predicting the gold standard of the training set images, the validation set is then used to optimise the model’s performance in predicting the gold standard of validation set images by optimising the hyperparameters of the model and the test set is used to relate the model’s overall performance in predicting the gold standard of a new dataset to which the model has not yet been exposed.

Finding the right trade-off between the relative shares of these three distinct datasets can be challenging. First, sufficient data are necessary to train a good model: the model should ‘see’ enough examples to ensure that the model can adequately predict the outcome of interest. Second, the validation set should be large enough to find the right hyperparameters and to prevent overfitting. Finally, the test set should be sufficiently large and heterogeneous to reflect the model’s performance in ‘real-life’, with enough statistical power for a meaningful estimation of performance metrics. This latter constraint is most important for to warrant the validity of the results and should never be compromised in favour of having more training examples. Ideally, multiple independent test sets are employed to assess the model’s robustness against different medical centres, acquisition protocols or patient ethnicities.

In addition, the balance between the respective class sizes should be considered carefully. Most papers report performance on a uniform class distribution (eg, 50% of cases with neoplastic lesions and 50% negative controls), while this is hardly ever the case in a real clinical setting, where the incidence of disease is generally much lower than 50%. Training with an approximately uniform class distribution is motivated by preventing the model to develop a bias towards either of the classes. However, researchers should report the expected class probabilities in a clinical setting and infer what their numbers from a uniform test setup would imply in the envisioned application. Table 4 shows an illustrative example of what happens for a system with a good performance in terms of sensitivity and specificity when the incidence drops, resulting a flood of spurious detections that can considerably dilute the number of true detections.

In most machine learning studies in endoscopy, data acquisition is more driven by data availability and convenience than by the optimal experimental design. This often leads to the use of a single database which is manually split into a training set and validation set, with the test set lacking or also originating from the same single database. Such a situation carries important limitations. First, because the three subsets used for training, validation and testing are derived from the same dataset, these subsets will be homogeneous, which will lead to good performance of the model when carried across the three sets. However, the overall dataset will likely suffer from selection bias and may not reflect the real-life exposure and natural variability. Second, the homogeneity of the three sets will not allow potential overfitting in the training phase or overfitting by hyperparameter optimisation in the validation phase to be picked up. The three datasets will have the same background noise, which will not directly be recognised but will be carried forward throughout the training-validation-test sets as relevant information for the model. Third, a manual split for creating a training set and an internal validation set from a single dataset may create subsets that are not independent. This holds especially if multiple images are derived from the same patient (eg, 800 images obtained from 50 patients) and subsets are created by dividing images instead of patients. In such instances, the training set and validation set will contain different images but stem from the same patient. Such images are clearly not independent, and the validation set will generate results that are too optimistic since they reflect overfitting. Therefore, the optimal experimental design of a machine learning model in endoscopy incorporates a training phase, a validation phase and a testing phase. Ideally, these should consist of independent datasets and not originate from a manual split of the same database.

Involved hyperparameters and method to determine their optimal values

The third crucial factor in the experimental design of machine learning algorithms relates to multiple testing of different models and cherry-picking of only the ‘best performing algorithms’. During the early phases of algorithm development, choices are made regarding the general model architectures, for example, one can use different CNNs, different ensemble learning techniques and different ways of cross-validation. While this is good practice at these early phases, during which it helps to select the most promising candidates, care should be taken not to redesign the architecture at later phases of algorithm development. When testing a large number of models, it is very likely that one model will yield good results on the test set. A common trend is that authors only select the best performing algorithms, and disregard the other models (ie, ‘cherry-picking’). A similar effect occurs when authors describe a cross-validation procedure, in which several models are trained and evaluated (one for each cross-validation fold), and subsequently continue their analysis only on the best performing fold. This is contradictory, since the rationale of cross-validation is to average out such outliers rather than to highlight them.