Session Item

Convolutional Neural Networks (CNN) can process imaging data with applications that span multiple medical disciplines. CNNs are considered to be less sensitive to data characteristics than traditional machine learning methods, however, the pre-processing of data can significantly impact the CNN performance. A common example of such pre-processing is applying a Hounsfield Unit (HU) window level and width to CT images to exclude artifacts and limit the range of tissues captured. In this work, we aim to explore the impact of windowing CT data for a CNN that predicts distant metastasis in head and neck (H&N) cancer patients.

We included 435 patients diagnosed with H&N cancer from two distinct publicly available datasets. One dataset, comprising 298 patients from 4 different institutions, was used for training and validation, leaving the other dataset, from a cohort with 137 patients, exclusively for testing. The gross tumor volume (GTV) was extracted from the pre-treatment CT scans using the mask delineations defined by an expert, cropped around the GTV region, resampled to a uniform pixel spacing (1 x 1 mm3), and windowed. For this last step, we considered different windowing parameters to evaluate the impact on the model’s performance. Firstly, a window based on a previous study on CNN models for H&N cancer outcome prediction, with a level of 0 Hounsfield units (HU) and a width of 1000 HU. Furthermore, a window with a level of 0 HU and a width of 500 HU to assess the effect of employing a narrower window. Lastly, we incorporated a window with a level of 125 HU and a width of 350 HU based on the expected interval of the Hounsfield scale for the tissues in the H&N region (e.g., mucosal, soft tissue). The CNN structure was adapted from a previous study and optimized for the prediction of 2-year distant metastasis. The model was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC) metric.

The model had the best performance with a window level of 125 HU and a width of 350 HU. This model achieved an average AUC of 0.88, 0.87, and 0.85 in the training, validation, and testing sets. The use of a wider window resulted in a lower AUC across the three subsets of data. Applying the windows with a width of 1000 HU and 500 HU resulted in an average AUC of 0.83 and 0.86 in the training set, 0.81 and 0.83 in the validation set, and 0.80 and 0.78 in the testing set. Significant differences (Kruskal-Wallis, p<0.05) were found for the validation and testing sets when comparing the AUCs of the best-performing window to the AUCs of models trained using the other two windows.

A CNN may attenuate differences in the pre-processing of imaging data due to its flexibility. However, pre-processing can still significantly impact a model's performance and generalizability. Mainly the windowing of CT scans using parameters based on the HUs of relevant tissues can positively impact the model's prediction ability.