Prediction of postoperative complications after oesophagectomy using machine-learning methods.

BACKGROUND: Oesophagectomy is an operation with a high risk of postoperative complications. The aim of this single-centre retrospective study was to apply machine-learning methods to predict complications (Clavien-Dindo grade IIIa or higher) and specific adverse events. METHODS: Patients with resectable adenocarcinoma or squamous cell carcinoma of the oesophagus and gastro-oesophageal junction who underwent Ivor Lewis oesophagectomy between 2016 and 2021 were included. The tested algorithms were logistic regression after recursive feature elimination, random forest, k-nearest neighbour, support vector machine, and neural network. The algorithms were also compared with a current risk score (the Cologne risk score). RESULTS: 457 patients had Clavien-Dindo grade IIIa or higher complications (52.9 per cent) versus 407 patients with Clavien-Dindo grade 0, I, or II complications (47.1 per cent). After 3-fold imputation and 3-fold cross-validation, the overall accuracies were: logistic regression after recursive feature elimination, 0.528; random forest, 0.535; k-nearest neighbour, 0.491; support vector machine, 0.511; neural network, 0.688; and Cologne risk score, 0.510. For medical complications, the results were: logistic regression after recursive feature elimination, 0.688; random forest, 0.664; k-nearest neighbour, 0.673; support vector machine, 0.681; neural network, 0.692; and Cologne risk score, 0.650. For surgical complications, the results were: logistic regression after recursive feature elimination, 0.621; random forest, 0.617; k-nearest neighbour, 0.620; support vector machine, 0.634; neural network, 0.667; and Cologne risk score, 0.624. The calculated area under the curve of the neural network was 0.672 for Clavien-Dindo grade IIIa or higher, 0.695 for medical complications, and 0.653 for surgical complications. CONCLUSION: The neural network scored the highest accuracies compared with all of the other models for the prediction of postoperative complications after oesophagectomy.

The human gullet or stomach can develop tumours. Surgery can help to cure patients with these tumours. But the operation is risky because sometimes adverse events can happen afterwards. So far, there is no reliable prediction model. It may help to predict the risk of adverse events accurately. For example, patients with a high risk could be observed more thoroughly. Patients with a low risk may not need unnecessary procedures. The information of all patients with an operation at a specialized hospital was collected. Machine learning is a complex mathematical method and was used in this study. It is able to analyse big data sets of information. One machine-learning method called neural network was best in predicting adverse events. Right now, the performance may not be strong enough to fully rely on the prediction. However, refinement of the prediction and more data could improve the neural network in the future.

Predicting the HER2 status in oesophageal cancer from tissue microarrays using convolutional neural networks.

BACKGROUND: Fast and accurate diagnostics are key for personalised medicine. Particularly in cancer, precise diagnosis is a prerequisite for targeted therapies, which can prolong lives. In this work, we focus on the automatic identification of gastroesophageal adenocarcinoma (GEA) patients that qualify for a personalised therapy targeting epidermal growth factor receptor 2 (HER2). We present a deep-learning method for scoring microscopy images of GEA for the presence of HER2 overexpression. METHODS: Our method is based on convolutional neural networks (CNNs) trained on a rich dataset of 1602 patient samples and tested on an independent set of 307 patient samples. We additionally verified the CNN's generalisation capabilities with an independent dataset with 653 samples from a separate clinical centre. We incorporated an attention mechanism in the network architecture to identify the tissue regions, which are important for the prediction outcome. Our solution allows for direct automated detection of HER2 in immunohistochemistry-stained tissue slides without the need for manual assessment and additional costly in situ hybridisation (ISH) tests. RESULTS: We show accuracy of 0.94, precision of 0.97, and recall of 0.95. Importantly, our approach offers accurate predictions in cases that pathologists cannot resolve and that require additional ISH testing. We confirmed our findings in an independent dataset collected in a different clinical centre. The attention-based CNN exploits morphological information in microscopy images and is superior to a predictive model based on the staining intensity only. CONCLUSIONS: We demonstrate that our approach not only automates an important diagnostic process for GEA patients but also paves the way for the discovery of new morphological features that were previously unknown for GEA pathology.

Noise reduction using novel loss functions to compute tissue mineral density and trabecular bone volume fraction on low resolution QCT.

Micro-structural parameters of the thoracic or lumbar spine generally carry insufficient accuracy and precision for clinical in vivo studies when assessed on quantitative computed tomography (QCT). We propose a 3D convolutional neural network with specific loss functions for QCT noise reduction to compute micro-structural parameters such as tissue mineral density (TMD) and bone volume ratio (BV/TV) with significantly higher accuracy than using no or standard noise reduction filters. The vertebra-phantom study contained high resolution peripheral and clinical CT scans with simulated in vivo CT noise and nine repetitions of three different tube currents (100, 250 and 360 mAs). Five-fold cross validation was performed on 20466 purely spongy pairs of noisy and ground-truth patches. Comparison of training and test errors revealed high robustness against over-fitting. While not showing effects for the assessment of BMD and voxel-wise densities, the filter improved thoroughly the computation of TMD and BV/TV with respect to the unfiltered data. Root-mean-square and accuracy errors of low resolution TMD and BV/TV decreased to less than 17% of the initial values. Furthermore filtered low resolution scans revealed still more TMD- and BV/TV-relevant information than high resolution CT scans, either unfiltered or filtered with two state-of-the-art standard denoising methods. The proposed architecture is threshold and rotational invariant, applicable on a wide range of image resolutions at once, and likely serves for an accurate computation of further micro-structural parameters. Furthermore, it is less prone for over-fitting than neural networks that compute structural parameters directly. In conclusion, the method is potentially important for the diagnosis of osteoporosis and other bone diseases since it allows to assess relevant 3D micro-structural information from standard low exposure CT protocols such as 100 mAs and 120 kVp.