Graph Structured Neural Networks for Perturbation Biology.

Computational modeling of perturbation biology identifies relationships between molecular elements and cellular response, and an accurate understanding of these systems will support the full realization of precision medicine. Traditional deep learning, while often accurate in predicting response, is unlikely to capture the true sequence of involved molecular interactions. Our work is motivated by two assumptions: 1) Methods that encourage mechanistic prediction logic are likely to be more trustworthy, and 2) problem-specific algorithms are likely to outperform generic algorithms. We present an alternative to Graph Neural Networks (GNNs) termed Graph Structured Neural Networks (GSNN), which uses cell signaling knowledge, encoded as a graph data structure, to add inductive biases to deep learning. We apply our method to perturbation biology using the LINCS L1000 dataset and literature-curated molecular interactions. We demonstrate that GSNNs outperform baseline algorithms in several prediction tasks, including 1) perturbed expression, 2) cell viability of drug combinations, and 3) disease-specific drug prioritization. We also present a method called GSNNExplainer to explain GSNN predictions in a biologically interpretable form. This work has broad application in basic biological research and pre-clincal drug repurposing. Further refinement of these methods may produce trustworthy models of drug response suitable for use as clinical decision aids. Availability and implementation: Our implementation of the GSNN method is available at https://github.com/nathanieljevans/GSNN. All data used in this work is publicly available.

Prediction of multiclass surgical outcomes in glaucoma using multimodal deep learning based on free-text operative notes and structured EHR data.

OBJECTIVE: Surgical outcome prediction is challenging but necessary for postoperative management. Current machine learning models utilize pre- and post-op data, excluding intraoperative information in surgical notes. Current models also usually predict binary outcomes even when surgeries have multiple outcomes that require different postoperative management. This study addresses these gaps by incorporating intraoperative information into multimodal models for multiclass glaucoma surgery outcome prediction. MATERIALS AND METHODS: We developed and evaluated multimodal deep learning models for multiclass glaucoma trabeculectomy surgery outcomes using both structured EHR data and free-text operative notes. We compare those to baseline models that use structured EHR data exclusively, or neural network models that leverage only operative notes. RESULTS: The multimodal neural network had the highest performance with a macro AUROC of 0.750 and F1 score of 0.583. It outperformed the baseline machine learning model with structured EHR data alone (macro AUROC of 0.712 and F1 score of 0.486). Additionally, the multimodal model achieved the highest recall (0.692) for hypotony surgical failure, while the surgical success group had the highest precision (0.884) and F1 score (0.775). DISCUSSION: This study shows that operative notes are an important source of predictive information. The multimodal predictive model combining perioperative notes and structured pre- and post-op EHR data outperformed other models. Multiclass surgical outcome prediction can provide valuable insights for clinical decision-making. CONCLUSIONS: Our results show the potential of deep learning models to enhance clinical decision-making for postoperative management. They can be applied to other specialties to improve surgical outcome predictions.

Towards Generalizability and Robustness in Biological Object Detection in Electron Microscopy Images.

Machine learning approaches have the potential for meaningful impact in the biomedical field. However, there are often challenges unique to biomedical data that prohibits the adoption of these innovations. For example, limited data, data volatility, and data shifts all compromise model robustness and generalizability. Without proper tuning and data management, deploying machine learning models in the presence of unaccounted for corruptions leads to reduced or misleading performance. This study explores techniques to enhance model generalizability through iterative adjustments. Specifically, we investigate a detection tasks using electron microscopy images and compare models trained with different normalization and augmentation techniques. We found that models trained with Group Normalization or texture data augmentation outperform other normalization techniques and classical data augmentation, enabling them to learn more generalized features. These improvements persist even when models are trained and tested on disjoint datasets acquired through diverse data acquisition protocols. Results hold true for transformerand convolution-based detection architectures. The experiments show an impressive 29% boost in average precision, indicating significant enhancements in the model's generalizibality. This underscores the models' capacity to effectively adapt to diverse datasets and demonstrates their increased resilience in real-world applications.

Efficient semi-supervised semantic segmentation of electron microscopy cancer images with sparse annotations.

Electron microscopy (EM) enables imaging at nanometer resolution and can shed light on how cancer evolves to develop resistance to therapy. Acquiring these images has become a routine task; however, analyzing them is now the bottleneck, as manual structure identification is very time-consuming and can take up to several months for a single sample. Deep learning approaches offer a suitable solution to speed up the analysis. In this work, we present a study of several state-of-the-art deep learning models for the task of segmenting nuclei and nucleoli in volumes from tumor biopsies. We compared previous results obtained with the ResUNet architecture to the more recent UNet++, FracTALResNet, SenFormer, and CEECNet models. In addition, we explored the utilization of unlabeled images through semi-supervised learning with Cross Pseudo Supervision. We have trained and evaluated all of the models on sparse manual labels from three fully annotated in-house datasets that we have made available on demand, demonstrating improvements in terms of 3D Dice score. From the analysis of these results, we drew conclusions on the relative gains of using more complex models, semi-supervised learning as well as next steps for the mitigation of the manual segmentation bottleneck.

Assessing the Impact of Image Quality on Deep Learning Classification of Infectious Keratitis.

Objective: To investigate the impact of corneal photograph quality on convolutional neural network (CNN) predictions. Design: A CNN trained to classify bacterial and fungal keratitis was evaluated using photographs of ulcers labeled according to 5 corneal image quality parameters: eccentric gaze direction, abnormal eyelid position, over/under-exposure, inadequate focus, and malpositioned light reflection. Participants: All eligible subjects with culture and stain-proven bacterial and/or fungal ulcers presenting to Aravind Eye Hospital in Madurai, India, between January 1, 2021 and December 31, 2021. Methods: Convolutional neural network classification performance was compared for each quality parameter, and gradient class activation heatmaps were generated to visualize regions of highest influence on CNN predictions. Main Outcome Measures: Area under the receiver operating characteristic and precision recall curves were calculated to quantify model performance. Bootstrapped confidence intervals were used for statistical comparisons. Logistic loss was calculated to measure individual prediction accuracy. Results: Individual presence of either light reflection or eyelids obscuring the corneal surface was associated with significantly higher CNN performance. No other quality parameter significantly influenced CNN performance. Qualitative review of gradient class activation heatmaps generally revealed the infiltrate as having the highest diagnostic relevance. Conclusions: The CNN demonstrated expert-level performance regardless of image quality. Future studies may investigate use of smartphone cameras and image sets with greater variance in image quality to further explore the influence of these parameters on model performance. Financial Disclosures: Proprietary or commercial disclosure may be found after the references.

Protein Phosphorylation Induced by Pyruvate Kinase M2 Inhibited Myofibrillar Protein Degradation in Post-Mortem Muscle.

Myofibrillar protein degradation is primarily related to meat tenderness through protein phosphorylation regulation. Pyruvate kinase M2 (PKM2), a glycolytic rate-limiting enzyme, is also regarded as a protein kinase to catalyze phosphorylation. The objective of this study was to investigate the relationship between myofibrillar protein degradation and phosphorylation induced by PKM2. Myofibrillar proteins were incubated with PKM2 at 4, 25, and 37 °C. The global phosphorylation level of myofibrillar proteins in the PKM2 group was significantly increased, but it was sensitive to temperature (P < 0.05). Compared with 4 and 25 °C, PKM2 significantly increased the myofibrillar protein phosphorylation level from 0.5 to 6 h at 37 °C (P < 0.05). In addition, the degradation of desmin and actin was inhibited after they were phosphorylated by PKM2 when incubated at 37 °C. These results demonstrate that phosphorylation of myofibrillar proteins catalyzed by PKM2 inhibited protein degradation and provided a possible pathway for meat tenderization through glycolytic enzyme regulation.

Automated large volume sample preparation for vEM.

New developments in electron microscopy technology, improved efficiency of detectors, and artificial intelligence applications for data analysis over the past decade have increased the use of volume electron microscopy (vEM) in the life sciences field. Moreover, sample preparation methods are continuously being modified by investigators to improve final sample quality, increase electron density, combine imaging technologies, and minimize the introduction of artifacts into specimens under study. There are a variety of conventional bench protocols that a researcher can utilize, though most of these protocols require several days. In this work, we describe the utilization of an automated specimen processor, the mPrep™ ASP-2000™, to prepare samples for vEM that are compatible with focused ion beam scanning electron microscopy (FIB-SEM), serial block face scanning electron microscopy (SBF-SEM), and array tomography (AT). The protocols described here aimed for methods that are completed in a much shorter period of time while minimizing the exposure of the operator to hazardous and toxic chemicals and improving the reproducibility of the specimens' heavy metal staining, all without compromising the quality of the data acquired using backscattered electrons during SEM imaging. As a control, we have included a widely used sample bench protocol and have utilized it as a comparator for image quality analysis, both qualitatively and using image quality analysis metrics.

Biology-inspired graph neural network encodes reactome and reveals biochemical reactions of disease.

Functional heterogeneity of healthy human tissues complicates interpretation of molecular studies, impeding precision therapeutic target identification and treatment. Considering this, we generated a graph neural network with Reactome-based architecture and trained it using 9,115 samples from Genotype-Tissue Expression (GTEx). Our graph neural network (GNN) achieves adjusted Rand index (ARI) = 0.7909, while a Resnet18 control model achieves ARI = 0.7781, on 370 held-out healthy human tissue samples from The Cancer Genome Atlas (TCGA), despite the Resnet18 using over 600 times the parameters. Our GNN also succeeds in separating 83 healthy skin samples from 95 lesional psoriasis samples, revealing that upregulation of 26S- and NUB1-mediated degradation of NEDD8, UBD, and their conjugates is central to the largest perturbed reaction network component in psoriasis. We show that our results are not discoverable using traditional differential expression and hypergeometric pathway enrichment analyses yet are supported by separate human multi-omics and small-molecule mouse studies, suggesting future molecular disease studies may benefit from similar GNN analytical approaches.

Prediction of Attention-Deficit/Hyperactivity Disorder Diagnosis Using Brief, Low-Cost Clinical Measures: A Competitive Model Evaluation.

Proper diagnosis of ADHD is costly, requiring in-depth evaluation via interview, multi-informant and observational assessment, and scrutiny of possible other conditions. The increasing availability of data may allow the development of machine-learning algorithms capable of accurate diagnostic predictions using low-cost measures to supplement human decision-making. We report on the performance of multiple classification methods used to predict a clinician-consensus ADHD diagnosis. Methods ranged from fairly simple (e.g., logistic regression) to more complex (e.g., random forest), while emphasizing a multi-stage Bayesian approach. Classifiers were evaluated in two large (N>1000), independent cohorts. The multi-stage Bayesian classifier provides an intuitive approach consistent with clinical workflows, and was able to predict expert consensus ADHD diagnosis with high accuracy (>86%)-though not significantly better than other methods. Results suggest that parent and teacher surveys are sufficient for high-confidence classifications in the vast majority of cases, while an important minority require additional evaluation for accurate diagnosis.

Deep Convolutional Neural Networks Detect no Morphological Differences Between Culture-Positive and Culture-Negative Infectious Keratitis Images.

Purpose: To determine whether convolutional neural networks can detect morphological differences between images of microbiologically positive and negative corneal ulcers. Methods: A cross-sectional comparison of prospectively collected data consisting of bacterial and fungal cultures and smears from eyes with acute infectious keratitis at Aravind Eye Hospital. Two convolutional neural network architectures (DenseNet and MobileNet) were trained using images obtained from handheld cameras collected from culture-positive and negative images and smear-positive and -negative images. Each architecture was trained on two image sets: (1) one with labels assigned using only culture results and (2) one using culture and smear results. The outcome measure was area under the receiver operating characteristic curve for predicting whether an ulcer would be microbiologically positive or negative. Results: There were 1970 images from 886 patients were included. None of the models were better than random chance at predicting positive microbiologic results (area under the receiver operating characteristic curve ranged from 0.49 to 0.56; all confidence intervals included 0.5). Conclusions: These two state-of-the-art deep convolutional neural network architectures could not reliably predict whether a corneal ulcer would be microbiologically positive or negative based on clinical photographs. This absence of detectable morphological differences informs the future development of computer vision models trained to predict the causative agent in infectious keratitis using corneal photography. Translational Relevance: These deep learning models were not able to identify morphological differences between microbiologically positive and negative corneal ulcers. This finding suggests that similar artificial intelligence models trained to identify the causative pathogen using only microbiologically positive cases may have potential to generalize well, including to cases with falsely negative microbiologic testing.

Emergency Department Pediatric Readiness Among US Trauma Centers: A Machine Learning Analysis of Components Associated With Survival.

OBJECTIVE: We used machine learning to identify the highest impact components of emergency department (ED) pediatric readiness for predicting in-hospital survival among children cared for in US trauma centers. BACKGROUND: ED pediatric readiness is associated with improved short-term and long-term survival among injured children and part of the national verification criteria for US trauma centers. However, the components of ED pediatric readiness most predictive of survival are unknown. METHODS: This was a retrospective cohort study of injured children below 18 years treated in 458 trauma centers from January 1, 2012, through December 31, 2017, matched to the 2013 National ED Pediatric Readiness Assessment and the American Hospital Association survey. We used machine learning to analyze 265 potential predictors of survival, including 152 ED readiness variables, 29 patient variables, and 84 ED-level and hospital-level variables. The primary outcome was in-hospital survival. RESULTS: There were 274,756 injured children, including 4585 (1.7%) who died. Nine ED pediatric readiness components were associated with the greatest increase in survival: policy for mental health care (+8.8% change in survival), policy for patient assessment (+7.5%), specific respiratory equipment (+7.2%), policy for reduced-dose radiation imaging (+7.0%), physician competency evaluations (+4.9%), recording weight in kilograms (+3.2%), life support courses for nursing (+1.0%-2.5%), and policy on pediatric triage (+2.5%). There was a 268% improvement in survival when the 5 highest impact components were present. CONCLUSIONS: ED pediatric readiness components related to specific policies, personnel, and equipment were the strongest predictors of pediatric survival and worked synergistically when combined.

Efficient semi-supervised semantic segmentation of electron microscopy cancer images with sparse annotations.

Electron microscopy (EM) enables imaging at a resolution of nanometers and can shed light on how cancer evolves to develop resistance to therapy. Acquiring these images has become a routine task.However, analyzing them is now a bottleneck, as manual structure identification is very time-consuming and can take up to several months for a single sample. Deep learning approaches offer a suitable solution to speed up the analysis. In this work, we present a study of several state-of-the-art deep learning models for the task of segmenting nuclei and nucleoli in volumes from tumor biopsies. We compared previous results obtained with the ResUNet architecture to the more recent UNet++, FracTALResNet, SenFormer, and CEECNet models. In addition, we explored the utilization of unlabeled images through semi-supervised learning with Cross Pseudo Supervision. We have trained and evaluated all of the models on sparse manual labels from three fully annotated in-house datasets that we have made available on demand, demonstrating improvements in terms of 3D Dice score. From the analysis of these results, we drew conclusions on the relative gains of using more complex models, and semi-supervised learning as well as the next steps for the mitigation of the manual segmentation bottleneck.

Segmentation of cellular ultrastructures on sparsely labeled 3D electron microscopy images using deep learning.

Focused ion beam-scanning electron microscopy (FIB-SEM) images can provide a detailed view of the cellular ultrastructure of tumor cells. A deeper understanding of their organization and interactions can shed light on cancer mechanisms and progression. However, the bottleneck in the analysis is the delineation of the cellular structures to enable quantitative measurements and analysis. We mitigated this limitation using deep learning to segment cells and subcellular ultrastructure in 3D FIB-SEM images of tumor biopsies obtained from patients with metastatic breast and pancreatic cancers. The ultrastructures, such as nuclei, nucleoli, mitochondria, endosomes, and lysosomes, are relatively better defined than their surroundings and can be segmented with high accuracy using a neural network trained with sparse manual labels. Cell segmentation, on the other hand, is much more challenging due to the lack of clear boundaries separating cells in the tissue. We adopted a multi-pronged approach combining detection, boundary propagation, and tracking for cell segmentation. Specifically, a neural network was employed to detect the intracellular space; optical flow was used to propagate cell boundaries across the z-stack from the nearest ground truth image in order to facilitate the separation of individual cells; finally, the filopodium-like protrusions were tracked to the main cells by calculating the intersection over union measure for all regions detected in consecutive images along z-stack and connecting regions with maximum overlap. The proposed cell segmentation methodology resulted in an average Dice score of 0.93. For nuclei, nucleoli, and mitochondria, the segmentation achieved Dice scores of 0.99, 0.98, and 0.86, respectively. The segmentation of FIB-SEM images will enable interpretative rendering and provide quantitative image features to be associated with relevant clinical variables.

Image-Based Differentiation of Bacterial and Fungal Keratitis Using Deep Convolutional Neural Networks.

Purpose: Develop computer vision models for image-based differentiation of bacterial and fungal corneal ulcers and compare their performance against human experts. Design: Cross-sectional comparison of diagnostic performance. Participants: Patients with acute, culture-proven bacterial or fungal keratitis from 4 centers in South India. Methods: Five convolutional neural networks (CNNs) were trained using images from handheld cameras collected from patients with culture-proven corneal ulcers in South India recruited as part of clinical trials conducted between 2006 and 2015. Their performance was evaluated on 2 hold-out test sets (1 single center and 1 multicenter) from South India. Twelve local expert cornea specialists performed remote interpretation of the images in the multicenter test set to enable direct comparison against CNN performance. Main Outcome Measures: Area under the receiver operating characteristic curve (AUC) individually and for each group collectively (i.e., CNN ensemble and human ensemble). Results: The best-performing CNN architecture was MobileNet, which attained an AUC of 0.86 on the single-center test set (other CNNs range, 0.68-0.84) and 0.83 on the multicenter test set (other CNNs range, 0.75-0.83). Expert human AUCs on the multicenter test set ranged from 0.42 to 0.79. The CNN ensemble achieved a statistically significantly higher AUC (0.84) than the human ensemble (0.76; P < 0.01). CNNs showed relatively higher accuracy for fungal (81%) versus bacterial (75%) ulcers, whereas humans showed relatively higher accuracy for bacterial (88%) versus fungal (56%) ulcers. An ensemble of the best-performing CNN and best-performing human achieved the highest AUC of 0.87, although this was not statistically significantly higher than the best CNN (0.83; P = 0.17) or best human (0.79; P = 0.09). Conclusions: Computer vision models achieved superhuman performance in identifying the underlying infectious cause of corneal ulcers compared with cornea specialists. The best-performing model, MobileNet, attained an AUC of 0.83 to 0.86 without any additional clinical or historical information. These findings suggest the potential for future implementation of these models to enable earlier directed antimicrobial therapy in the management of infectious keratitis, which may improve visual outcomes. Additional studies are ongoing to incorporate clinical history and expert opinion into predictive models.

Malignancy Detection in Prostate Multi-Parametric MR Images Using U-net with Attention.

Multiparametric magnetic resonance (mpMR) images are increasingly being used for diagnosis and monitoring of prostate cancer. Detection of malignancy from prostate mpMR images requires expertise, is time consuming and prone to human error. The recent developments of U-net have demonstrated promising detection results in many medical applications. Straightforward use of U-net tends to result in over-detection in mpMR images. The recently developed attention mechanism can help retain only features relevant for malignancy detection, thus improving the detection accuracy. In this work, we propose a U-net architecture that is enhanced by the attention mechanism to detect malignancy in prostate mpMR images. This approach resulted in improved performance in terms of higher Dice score and reduced over-detection when compared to U-net in detecting malignancy.

Early Prediction of Breast Cancer Therapy Response using Multiresolution Fractal Analysis of DCE-MRI Parametric Maps.

We aimed to determine whether multiresolution fractal analysis of voxel-based dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) parametric maps can provide early prediction of breast cancer response to neoadjuvant chemotherapy (NACT). In total, 55 patients underwent 4 DCE-MRI examinations before, during, and after NACT. The shutter-speed model was used to analyze the DCE-MRI data and generate parametric maps within the tumor region of interest. The proposed multiresolution fractal method and the more conventional methods of single-resolution fractal, gray-level co-occurrence matrix, and run-length matrix were used to extract features from the parametric maps. Only the data obtained before and after the first NACT cycle were used to evaluate early prediction of response. With a training (N = 40) and testing (N = 15) data set, support vector machine was used to assess the predictive abilities of the features in classification of pathologic complete response versus non-pathologic complete response. Generally the multiresolution fractal features from individual maps and the concatenated features from all parametric maps showed better predictive performances than conventional features, with receiver operating curve area under the curve (AUC) values of 0.91 (all parameters) and 0.80 (Ktrans), in the training and testing sets, respectively. The differences in AUC were statistically significant (P < .05) for several parametric maps. Thus, multiresolution analysis that decomposes the texture at various spatial-frequency scales may more accurately capture changes in tumor vascular heterogeneity as measured by DCE-MRI, and therefore provide better early prediction of NACT response.

Performance of Microbiome Sequence Inference Methods in Environments with Varying Biomass.

Microbiome community composition plays an important role in human health, and while most research to date has focused on high-microbial-biomass communities, low-biomass communities are also important. However, contamination and technical noise make determining the true community signal difficult when biomass levels are low, and the influence of varying biomass on sequence processing methods has received little attention. Here, we benchmarked six methods that infer community composition from 16S rRNA sequence reads, using samples of varying biomass. We included two operational taxonomic unit (OTU) clustering algorithms, one entropy-based method, and three more-recent amplicon sequence variant (ASV) methods. We first compared inference results from high-biomass mock communities to assess baseline performance. We then benchmarked the methods on a dilution series made from a single mock community-samples that varied only in biomass. ASVs/OTUs inferred by each method were classified as representing expected community, technical noise, or contamination. With the high-biomass data, we found that the ASV methods had good sensitivity and precision, whereas the other methods suffered in one area or in both. Inferred contamination was present only in small proportions. With the dilution series, contamination represented an increasing proportion of the data from the inferred communities, regardless of the inference method used. However, correlation between inferred contaminants and sample biomass was strongest for the ASV methods and weakest for the OTU methods. Thus, no inference method on its own can distinguish true community sequences from contaminant sequences, but ASV methods provide the most accurate characterization of community and contaminants. IMPORTANCE Microbial communities have important ramifications for human health, but determining their impact requires accurate characterization. Current technology makes microbiome sequence data more accessible than ever. However, popular software methods for analyzing these data are based on algorithms developed alongside older sequencing technology and smaller data sets and thus may not be adequate for modern, high-throughput data sets. Additionally, samples from environments where microbes are scarce present additional challenges to community characterization relative to high-biomass environments, an issue that is often ignored. We found that a new class of microbiome sequence processing tools, called amplicon sequence variant (ASV) methods, outperformed conventional methods. In samples representing low-biomass communities, where sample contamination becomes a significant confounding factor, the improved accuracy of ASV methods may allow more-robust computational identification of contaminants.

Growth of graphene with large single-crystal domains by Ni foam-assisted structure and its high-gain field-effect transistors.

High-quality graphene materials and high-performance graphene transistors have attracted much attention in recent years. To obtain high-performance graphene transistors, large single-crystal graphene is needed. The synthesis of large-domain-sized single-crystal graphene requires low nucleation density; this can lead to a lower growth rate. In this study, a Ni-foam assisted structure was developed to control the nucleation density and growth rate of graphene by tuning the flow dynamics. Lower nucleation density and high growth rate (∼50 µm min-1) were achieved with a 4 mm-gap Ni foam. With the graphene transistor fabrication process, a pre-deposited Au film as the protective layer was used during the graphene transfer. Graphene transistors showed good current saturation with drain differential conductance as low as 0.04 S mm-1 in the strong saturation region. For the devices with gate length of 2 µm, the intrinsic cut-off frequency f T and maximum oscillation frequency f max were 8.4 and 16.3 GHz, respectively, with f max/f T = 1.9 and power gain of up to 6.4 dB at 1 GHz. The electron velocity saturation induced by the surface optical phonons of SiO2 substrates was analyzed. Electron velocity saturation and ultra-thin Al2O3 gate dielectrics were thought to be the reasons for the good current saturation and high power gain of the graphene transistors.

Analysis of DCE-MRI for Early Prediction of Breast Cancer Therapy Response.

Positive response to neoadjuvant chemotherapy (NACT) has been correlated to better long-term outcomes in breast cancer treatment. Early prediction of response to NACT can help modify the regimen for non-responding patients, sparing them of potential toxicities of ineffective therapies. It has been observed that tumor functions such as vascularization and vascular permeability change even before noticeable changes occur in the tumor size in response to the treatment. Therefore, it is essential to have reliable imaging based features to measure these changes. Texture analysis on parametric maps from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) has shown to be a good predictor of breast cancer response to NACT at an early stage. But hand crafted texture features might not be able to capture the rich spatio-temporal information in the parametric maps. In this work, we studied the ability of convolutional neural networks in predicting the response to NACT at an early stage.