Scientists from the York University, Toronto, Canada, investigated a novel deep learning-based methodology to predict breast cancer response to neo-adjuvant chemotherapy (NAC) using the quantitative ultrasound (QUS) multiparametric imaging at pre-treatment.
Breast cancer is the most general type of cancer in women and the leading cause of cancer-related death. Locally advanced breast cancer (LABC) is an invasive breast cancer subtype that accounts for up to 20% of new occurrences each year. LABC is commonly diagnosed in cancers larger than 5 cm in diameter, as well as skin and/or chest wall involvement. Patients with metastatic breast cancer or numerous positive axillary lymph nodes are also included in LABC.
Because of the variety of systemic and targeted regimens available, NAC followed by surgery is now the usual treatment for LABC patients. In some circumstances, adjuvant radiation and/or hormone therapy are used after surgery to lower the risk of cancer recurrence. The conventional way to determine the tumor pathological response to NAC is post-surgical histology. Post-surgical evaluations, on the other hand, cannot be utilized to alter the NAC or transition to salvage treatment. Early detection of tumor response to NAC allows for therapeutic changes such as changing the treatment regimen, dose, and/or sequence, as well as moving to more effective treatments or even salvage therapies before it’s too late for particular patients.
Quantitative ultrasound (QUS) approaches have been developed to produce quantitative estimates of tissue biophysical parameters that are unrelated to instrument settings and need less operator involvement. QUS spectral characteristics have been found to detect tumor cell death triggered by various anticancer therapy. Moreover, several studies have shown that hand-crafted features derived from QUS parametric maps can also be used to anticipate and supervise breast cancer reaction to neoadjuvant chemotherapy before or even within weeks of treatment beginning, with high correlations to clinical and pathological response recognized at the end of therapy.
Deep Learning methodologies have the potential to eliminate the need for traditional machine learning techniques to extract carefully created handmade features from images. Deep learning frameworks optimize their data-driven feature maps during the iterative training process.
In this study, the efficiency of deep convolutional neural networks (DCNN) techniques on QUS spectral multi-parametric pictures to predict LABC response to NAC before the start of therapy was explored.
A total of 181 eligible patients were enrolled in the research, with each receiving a core needle biopsy to confirm the cancer diagnosis and grade the tumor. About 30 percent (n=50) of the 181 patients were randomly picked using stratified random selection and kept unobserved as an independent test set, while the other patients (n=131) were utilized to create and optimize the predictive models.
The patients were divided into two groups: responders and non-responders, using a modified response (MR) grading method based on response assessment criteria in solid tumors (RECIST) and histological criteria. Patients with an MR score of 1–2 (under 30% tumor reduction) were classified as non-responders, whereas those with an MR score of 3–5 (greater than 30% tumor reduction or extremely low residual tumor cellularity) were classified as responders. In accordance with this, responders and non-responders were identified as 138 and 43 patients, respectively.
The ultrasound data for each tumor was collected in four to seven imaging planes across the breast at roughly 1 cm intervals. The image size was 6 cm in the lateral and 4–6 cm in the axial directions, respectively.
Using a sliding window analysis, parametric maps were created for all imaging planes of the tumor throughout the full area of interest (tumor core and margin).
Two cascaded networks make up the deep learning framework. The feature network in this research is a DCNN with many convolutional layers as its backbone that has been tuned to extract the best feature maps from QUS parametric photos. The backbone of the feature network was explored in this study using two primary architectures: a modified residual network version 101 (ResNet) and a modified residual attention network version 56 (RAN). The feature network’s optimized feature maps are averaged across all parametric pictures associated with each tumor and then employed in a second fully connected network (predictive network) customized for patient response prediction. To anticipate the probability of the response categories (responder vs. non-responder) for each patient, the predictive network comprises two fully-connected layers with an input layer, a middle layer, and a softmax layer at the end. To avoid overfitting and improve generalization performance, drop-out layers have been inserted after the first and second layers of this network.
The parametric images were pre-processed and modified for the convolutional model before training the models. To optimize the network hyperparameters, about 25% of the training set (31 patients) was chosen at random as a validation set. The QUS multi-parametric pictures (MBF, SI, ESD, and EAC) of the tumor core, as well as the core and its margin, were evaluated as inputs to the DCNN framework in several trials and their performance in response prediction was compared. The training set was used to train and optimize deep learning models with various feature networks. The accuracy, sensitivity, specificity, and ROC analysis were used to evaluate the performance of the optimized models on an independent test set.
The patients’ initial tumor size was 5.2 cm on average, with a 2.5 cm residual tumor size at the end of treatment. At the treatment completion, 76.2 percent and 23.8 percent of the patients were classified as responders and non-responders, respectively, using the MR grading system. Minimal tumor cellularity frequently remained inside the tumor bed after treatment in responding individuals, as seen by histopathology slides. Non-responding patients’ histopathological pictures, on the other hand, generally showed vast regions of residual illness with no chemotherapeutic effects. On the independent test set, using the ResNet architecture as the model’s backbone to extract feature maps from parametric pictures of the tumor core resulted in an AUC of 0.77. The AUC of this model was increased to 0.83 by expanding the input parametric images incorporating both the tumor core and its margin.
Using the RAN architecture to apply the extracted features from the parametric images of the tumor core resulted in an accuracy of 80% and an AUC of 0.82 on the independent test set. The overall performance of this model was improved by including the tumor margin in the input parametric images. On the independent test set, this model had the best prediction performance, with accuracy and AUC of 88 % and 0.86, respectively. Long-term survival assessments revealed statistically significant differences in the survival of responders and non-responders detected at pre-treatment using this model’s prediction, as well as at post-treatment using traditional clinical and pathological parameters.
Future diversified analyses on larger datasets, including distinct models with different NAC regimes and possibly different molecular subtypes, could lead to more powerful predictive models. Future randomized clinical trials to evaluate the adjustment of the NAC regimen for individuals with a poor likelihood of response to standard treatment could benefit from such models. Finally, this study showed that DCNN models could be used to predict therapy response in the context of quantitative imaging. The attention-guided convolutional networks performed better in generating optimal quantitative features from QUS multi-parametric pictures, according to the findings.
The deep learning models constructed in this work had a high degree of accuracy in predicting the survival-related response of LABC patients to NAC before they began treatment. Predicting the outcome of NAC treatment from the start would allow clinicians to make adjustments to ineffective treatment protocols for specific patients. The findings of this work are encouraging, prompting further research employing various DCNN designs and data from bigger (multi-institutional) patient cohorts to assess the robustness of the techniques in the clinic.
Story Source: Taleghamar, H., Jalalifar, S.A., Czarnota, G.J. et al. Deep learning of quantitative ultrasound multi-parametric images at pre-treatment to predict breast cancer response to chemotherapy. Sci Rep 12, 2244 (2022)
Data Availability: https://doi.org/10.1038/s41598-022-06100-2