Item 22 — Detailed description of model

If novel model architecture is used, provide a complete and detailed structure of the model, including inputs, outputs, and all intermediate layers, in sufficient detail that another investigator could exactly reconstruct the network. For neural network models, include all details of pooling, normalization, regularization, and activation in the layer descriptions. Model inputs must match the form of the preprocessed data. Model outputs must correspond to the requirements of the stated clinical problem, and for supervised learning should match the form of the reference standard annotations. If a previously published model architecture is employed, cite a reference that meets the preceding standards and fully describe every modification made to the model. Cite a reference for any proprietary model described previously, as well. In some cases, it may be more convenient to provide the structure of the model in code as supplemental data.

Explanation and Elaboration: A full description of model architecture is essential to ensure that other researchers can replicate the model or build upon the study. Checking that model inputs match what was described for preprocessing steps helps to identify steps that may have been performed but not described. For instance, if a network operating on radiographs takes a fixed input size of 224 x 224 pixels but there is no downscaling step described in the preprocessing, something is clearly missing. Similarly, if the output of the model is not directly applicable to the stated clinical problem either the model will not be clinically useful or there are post-processing steps that must be described. Large models may be difficult to succinctly but completely describe in prose, and it may not be apparent when there are missing details. Code that constructs the model using a widely used framework may be a more direct and useful means of fully specifying the model than a natural language (e.g. English) description.

Examples
Specifically, a DL-based segmentation model was developed for each target structure by using a network based on encoder-decoder architecture, that is, ResNet-50–U-Net …. The encoder ResNet-50 extracted features from the original image, and the decoder U-Net completed the fusion of information based on these features, thereby completing the pixel-by-pixel prediction of input image. In more detail, ResNet-50 is a 50-layer convolutional neural network that uses skip connections to propagate information over layers, allowing for increased depth of the network and higher performance levels compared with other models …. The ResNet-50 was pretrained on the ImageNet 2012 classification dataset (https://www.image-net.org/), which comprises 1000 classes and 1.28 million training images. The pretrained weights were obtained from a public repository (https://github.com/fchollet/deep-learning-models/releases/download/v0.2/resnet50_weights_tf_dim_ordering_tf_kernels.h5). The training of each DL model was performed in five epochs using the Adam optimizer with an initial learning rate of 0.01. Moreover, data augmentation was used to increase the size of the training dataset to enhance the performance of DL models. Specifically, data augmentation was achieved by rotating (−30° < angle θ < 30°) and translating images by −20% to 20% on the x- and y-axis independently. [Materials and methods (36), CC BY 4.0]
Classification ensemble constituents used a modified InceptionV3 architecture …. Input layers were adjusted to accept 764 × 764-pixel and 1024 × 1024-pixel resolution DICOM files. The classification head consisted of a global average pooling layer, followed by a dense layer with rectified linear unit … activation and a dropout layer. A softmax activation function was applied to the final dense layer. Model ensemble output probabilities were calculated by averaging probabilities of each constituent model. The number of neurons in the penultimate dense layer and the dropout rate were tunable hyperparameters optimized during training using the Hyperband algorithm …, with the best set of parameters corresponding to the lowest categorical cross-entropy loss on the validation set. The number of neurons was selected from the range of [32, 512], and the dropout rate took values from the range [0, 0.2]. Models were trained in two phases. In the first phase (pretraining), networks were initialized with ImageNet weights … and trained to differentiate normal and abnormal chest radiographs using transformer-generated ground truths from 1 132 142 DICOM images and linked radiology reports (Fig 1). In the second phase (classification), models were initialized with the pretrained weights and further fine-tuned on manually annotated NGT DICOM images. Training was performed with 32 images per batch using an Adam optimizer with a learning rate of 1 × 10−3 while minimizing the categorical cross-entropy loss. Input images were resized to 764 × 764 or 1024 × 1024 pixels using bilinear interpolation without preserving the aspect ratio. During training, images were subject to random augmentations, which included brightness adjustments, angular rotation, and left-to-right flipping. Training was terminated early if validation loss did not improve after 10 consecutive epochs. [Materials and methods (50), CC BY 4.0]