Explainable 3D CNN based on baseline breast DCE-MRI to give an early prediction of pathological complete response to neoadjuvant chemotherapy.

BACKGROUND: So far, baseline Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) has played a key role for the application of sophisticated artificial intelligence-based models using Convolutional Neural Networks (CNNs) to extract quantitative imaging information as earlier indicators of pathological Complete Response (pCR) achievement in breast cancer patients treated with neoadjuvant chemotherapy (NAC). However, these models did not exploit the DCE-MRI exams in their full geometry as 3D volume but analysed only few individual slices independently, thus neglecting the depth information. METHOD: This study aimed to develop an explainable 3D CNN, which fulfilled the task of pCR prediction before the beginning of NAC, by leveraging the 3D information of post-contrast baseline breast DCE-MRI exams. Specifically, for each patient, the network took in input a 3D sequence containing the tumor region, which was previously automatically identified along the DCE-MRI exam. A visual explanation of the decision-making process of the network was also provided. RESULTS: To the best of our knowledge, our proposal is competitive than other models in the field, which made use of imaging data alone, reaching a median AUC value of 81.8%, 95%CI [75.3%; 88.3%], a median accuracy value of 78.7%, 95%CI [74.8%; 82.5%], a median sensitivity value of 69.8%, 95%CI [59.6%; 79.9%] and a median specificity value of 83.3%, 95%CI [82.6%; 84.0%], respectively. The median and CIs were computed according to a 10-fold cross-validation scheme for 5 rounds. CONCLUSION: Finally, this proposal holds high potential to support clinicians on non-invasively early pursuing or changing patient-centric NAC pathways.

Comparison between vision transformers and convolutional neural networks to predict non-small lung cancer recurrence.

Non-Small cell lung cancer (NSCLC) is one of the most dangerous cancers, with 85% of all new lung cancer diagnoses and a 30-55% of recurrence rate after surgery. Thus, an accurate prediction of recurrence risk in NSCLC patients during diagnosis could be essential to drive targeted therapies preventing either overtreatment or undertreatment of cancer patients. The radiomic analysis of CT images has already shown great potential in solving this task; specifically, Convolutional Neural Networks (CNNs) have already been proposed providing good performances. Recently, Vision Transformers (ViTs) have been introduced, reaching comparable and even better performances than traditional CNNs in image classification. The aim of the proposed paper was to compare the performances of different state-of-the-art deep learning algorithms to predict cancer recurrence in NSCLC patients. In this work, using a public database of 144 patients, we implemented a transfer learning approach, involving different Transformers architectures like pre-trained ViTs, pre-trained Pyramid Vision Transformers, and pre-trained Swin Transformers to predict the recurrence of NSCLC patients from CT images, comparing their performances with state-of-the-art CNNs. Although, the best performances in this study are reached via CNNs with AUC, Accuracy, Sensitivity, Specificity, and Precision equal to 0.91, 0.89, 0.85, 0.90, and 0.78, respectively, Transformer architectures reach comparable ones with AUC, Accuracy, Sensitivity, Specificity, and Precision equal to 0.90, 0.86, 0.81, 0.89, and 0.75, respectively. Based on our preliminary experimental results, it appears that Transformers architectures do not add improvements in terms of predictive performance to the addressed problem.

Prognostic power assessment of clinical parameters to predict neoadjuvant response therapy in HER2-positive breast cancer patients: A machine learning approach.

BACKGROUND: About 15%-20% of breast cancer (BC) cases is classified as Human Epidermal growth factor Receptor type 2 (HER2) positive. The Neoadjuvant chemotherapy (NAC) was initially introduced for locally advanced and inflammatory BC patients to allow a less extensive surgical resection, whereas now it represents the current standard for early-stage and operable BC. However, only 20%-40% of patients achieve pathologic complete response (pCR). According to the results of practice-changing clinical trials, the addition of trastuzumab to NAC brings improvements to pCR, and recently, the use of pertuzumab plus trastuzumab has registered further statistically significant and clinically meaningful improvements in terms of pCR. The goal of our work is to propose a machine learning model to predict the pCR to NAC in HER2-positive patients based on a subset of clinical features. METHOD: First, we evaluated the significant association of clinical features with pCR on the retrospectively collected data referred to 67 patients afferent to Istituto Tumori "Giovanni Paolo II." Then, we performed a feature selection procedure to identify a subset of features to be used for training a machine learning-based classification algorithm. As a result, pCR to NAC was associated with ER status, Pgr status, and HER2 score. RESULTS: The machine learning model trained on a subgroup of essential features reached an AUC of 73.27% (72.44%-73.66%) and an accuracy of 71.67% (71.64%-73.13%). According to our results, the clinical features alone are not enough to define a support system useful for clinical pathway. CONCLUSION: Our results seem worthy of further investigation in large validation studies and this work could be the basis of future study that will also involve radiomics analysis of biomedical images.

An explainable machine learning ensemble model to predict the risk of ovarian cancer in BRCA-mutated patients undergoing risk-reducing salpingo-oophorectomy.

Introduction: It has been estimated that 19,880 new cases of ovarian cancer had been diagnosed in 2022. Most epithelial ovarian cancer are sporadic, while in 15%-25% of cases, there is evidence of a familial or inherited component. Approximately 20%-25% of high-grade serous carcinoma cases are caused by germline mutations in the BRCA1 and BRCA2 genes. However, owing to a lack of effective early detection methods, women with BRCA mutations are recommended to undergo bilateral risk-reducing salpingo-oophorectomy (RRSO) after childbearing. Determining the right timing for this procedure is a difficult decision. It is crucial to find a clinical signature to identify high-risk BRCA-mutated patients and determine the appropriate timing for performing RRSO. Methods: In this work, clinical data referred to a cohort of 184 patients, of whom 7.6% were affected by adnexal tumors including invasive carcinomas and intraepithelial lesions after RSSO has been analyzed. Thus, we proposed an explainable machine learning (ML) ensemble approach using clinical data commonly collected in clinical practice to early identify BRCA-mutated patients at high risk of ovarian cancer and consequentially establish the correct timing for RRSO. Results: The ensemble model was able to handle imbalanced data achieving an accuracy value of 83.2%, a specificity value of 85.3%, a sensitivity value of 57.1%, a G-mean value of 69.8%, and an AUC value of 71.1%. Discussion: In agreement with the promising results achieved, the application of suitable ML techniques could play a key role in the definition of a BRCA-mutated patient-centric clinical signature for ovarian cancer risk and consequently personalize the management of these patients. As far as we know, this is the first work addressing this task from an ML perspective.

A CT-based transfer learning approach to predict NSCLC recurrence: The added-value of peritumoral region.

Non-small cell lung cancer (NSCLC) represents 85% of all new lung cancer diagnoses and presents a high recurrence rate after surgery. Thus, an accurate prediction of recurrence risk in NSCLC patients at diagnosis could be essential to designate risk patients to more aggressive medical treatments. In this manuscript, we apply a transfer learning approach to predict recurrence in NSCLC patients, exploiting only data acquired during its screening phase. Particularly, we used a public radiogenomic dataset of NSCLC patients having a primary tumor CT image and clinical information. Starting from the CT slice containing the tumor with maximum area, we considered three different dilatation sizes to identify three Regions of Interest (ROIs): CROP (without dilation), CROP 10 and CROP 20. Then, from each ROI, we extracted radiomic features by means of different pre-trained CNNs. The latter have been combined with clinical information; thus, we trained a Support Vector Machine classifier to predict the NSCLC recurrence. The classification performances of the devised models were finally evaluated on both the hold-out training and hold-out test sets, in which the original sample has been previously divided. The experimental results showed that the model obtained analyzing CROP 20 images, which are the ROIs containing more peritumoral area, achieved the best performances on both the hold-out training set, with an AUC of 0.73, an Accuracy of 0.61, a Sensitivity of 0.63, and a Specificity of 0.60, and on the hold-out test set, with an AUC value of 0.83, an Accuracy value of 0.79, a Sensitivity value of 0.80, and a Specificity value of 0.78. The proposed model represents a promising procedure for early predicting recurrence risk in NSCLC patients.

Assessing the cost-effectiveness of waiting list reduction strategies for a breast radiology department: a real-life case study.

BACKGROUND: A timely diagnosis is essential for improving breast cancer patients' survival and designing targeted therapeutic plans. For this purpose, the screening timing, as well as the related waiting lists, are decisive. Nonetheless, even in economically advanced countries, breast cancer radiology centres fail in providing effective screening programs. Actually, a careful hospital governance should encourage waiting lists reduction programs, not only for improving patients care, but also for minimizing costs associated with the treatment of advanced cancers. Thus, in this work, we proposed a model to evaluate several scenarios for an optimal distribution of the resources invested in a Department of Breast Radiodiagnosis. MATERIALS AND METHODS: Particularly, we performed a cost-benefit analysis as a technology assessment method to estimate both costs and health effects of the screening program, to maximise both benefits related to the quality of care and resources employed by the Department of Breast Radiodiagnosis of Istituto Tumori "Giovanni Paolo II" of Bari in 2019. Specifically, we determined the Quality-Adjusted Life Year (QALY) for estimating health outcomes, in terms of usefulness of two hypothetical screening strategies with respect to the current one. While the first hypothetical strategy adds one team made up of a doctor, a technician and a nurse, along with an ultrasound and a mammograph, the second one adds two afternoon teams. RESULTS: This study showed that the most cost-effective incremental ratio could be achieved by reducing current waiting lists from 32 to 16 months. Finally, our analysis revealed that this strategy would also allow to include more people in the screening programs (60,000 patients in 3 years).

Machine learning survival models trained on clinical data to identify high risk patients with hormone responsive HER2 negative breast cancer.

For endocrine-positive Her2 negative breast cancer patients at an early stage, the benefit of adding chemotherapy to adjuvant endocrine therapy is not still confirmed. Several genomic tests are available on the market but are very expensive. Therefore, there is the urgent need to explore novel reliable and less expensive prognostic tools in this setting. In this paper, we shown a machine learning survival model to estimate Invasive Disease-Free Events trained on clinical and histological data commonly collected in clinical practice. We collected clinical and cytohistological outcomes of 145 patients referred to Istituto Tumori "Giovanni Paolo II". Three machine learning survival models are compared with the Cox proportional hazards regression according to time-dependent performance metrics evaluated in cross-validation. The c-index at 10 years obtained by random survival forest, gradient boosting, and component-wise gradient boosting is stabled with or without feature selection at approximately 0.68 in average respect to 0.57 obtained to Cox model. Moreover, machine learning survival models have accurately discriminated low- and high-risk patients, and so a large group which can be spared additional chemotherapy to hormone therapy. The preliminary results obtained by including only clinical determinants are encouraging. The integrated use of data already collected in clinical practice for routine diagnostic investigations, if properly analyzed, can reduce time and costs of the genomic tests.

A Machine Learning Approach for Predicting Capsular Contracture after Postmastectomy Radiotherapy in Breast Cancer Patients.

In recent years, immediate breast reconstruction after mastectomy surgery has steadily increased in the treatment pathway of breast cancer (BC) patients due to its potential impact on both the morpho-functional and aesthetic type of the breast and the quality of life. Although recent studies have demonstrated how recent radiotherapy techniques have allowed a reduction of adverse events related to breast reconstruction, capsular contracture (CC) remains the main complication after post-mastectomy radio-therapy (PMRT). In this study, we evaluated the association of the occurrence of CC with some clinical, histological and therapeutic parameters related to BC patients. We firstly performed bivariate statistical tests and we then evaluated the prognostic predictive power of the collected data by using machine learning techniques. Out of a sample of 59 patients referred to our institute, 28 patients (i.e., 47%) showed contracture after PMRT. As a result, only estrogen receptor status (ER) and molecular subtypes were significantly associated with the occurrence of CC after PMRT. Different machine learning models were trained on a subset of clinical features selected by a feature importance approach. Experimental results have shown that collected features have a non-negligible predictive power. The extreme gradient boosting classifier achieved an area under the curve (AUC) value of 68% and accuracy, sensitivity, and specificity values of 68%, 64%, and 74%, respectively. Such a support tool, after further suitable optimization and validation, would allow clinicians to identify the best therapeutic strategy and reconstructive timing.

Analyzing breast cancer invasive disease event classification through explainable artificial intelligence.

Introduction: Recently, accurate machine learning and deep learning approaches have been dedicated to the investigation of breast cancer invasive disease events (IDEs), such as recurrence, contralateral and second cancers. However, such approaches are poorly interpretable. Methods: Thus, we designed an Explainable Artificial Intelligence (XAI) framework to investigate IDEs within a cohort of 486 breast cancer patients enrolled at IRCCS Istituto Tumori "Giovanni Paolo II" in Bari, Italy. Using Shapley values, we determined the IDE driving features according to two periods, often adopted in clinical practice, of 5 and 10 years from the first tumor diagnosis. Results: Age, tumor diameter, surgery type, and multiplicity are predominant within the 5-year frame, while therapy-related features, including hormone, chemotherapy schemes and lymphovascular invasion, dominate the 10-year IDE prediction. Estrogen Receptor (ER), proliferation marker Ki67 and metastatic lymph nodes affect both frames. Discussion: Thus, our framework aims at shortening the distance between AI and clinical practice.

Corrigendum: Transfer learning approach based on computed tomography images for predicting late xerostomia after radiotherapy in patients with oropharyngeal cancer.

[This corrects the article DOI: 10.3389/fmed.2022.993395.].

A deep learning model based on whole slide images to predict disease-free survival in cutaneous melanoma patients.

The application of deep learning on whole-slide histological images (WSIs) can reveal insights for clinical and basic tumor science investigations. Finding quantitative imaging biomarkers from WSIs directly for the prediction of disease-free survival (DFS) in stage I-III melanoma patients is crucial to optimize patient management. In this study, we designed a deep learning-based model with the aim of learning prognostic biomarkers from WSIs to predict 1-year DFS in cutaneous melanoma patients. First, WSIs referred to a cohort of 43 patients (31 DF cases, 12 non-DF cases) from the Clinical Proteomic Tumor Analysis Consortium Cutaneous Melanoma (CPTAC-CM) public database were firstly annotated by our expert pathologists and then automatically split into crops, which were later employed to train and validate the proposed model using a fivefold cross-validation scheme for 5 rounds. Then, the model was further validated on WSIs related to an independent test, i.e. a validation cohort of 11 melanoma patients (8 DF cases, 3 non-DF cases), whose data were collected from Istituto Tumori 'Giovanni Paolo II' in Bari, Italy. The quantitative imaging biomarkers extracted by the proposed model showed prognostic power, achieving a median AUC value of 69.5% and a median accuracy of 72.7% on the public cohort of patients. These results remained comparable on the validation cohort of patients with an AUC value of 66.7% and an accuracy value of 72.7%, respectively. This work is contributing to the recently undertaken investigation on how treat features extracted from raw WSIs to fulfil prognostic tasks involving melanoma patients. The promising results make this study as a valuable basis for future research investigation on wider cohorts of patients referred to our Institute.

Transfer learning approach based on computed tomography images for predicting late xerostomia after radiotherapy in patients with oropharyngeal cancer.

Background and purpose: Although the latest breakthroughs in radiotherapy (RT) techniques have led to a decrease in adverse event rates, these techniques are still associated with substantial toxicity, including xerostomia. Imaging biomarkers could be useful to predict the toxicity risk related to each individual patient. Our preliminary work aims to develop a radiomic-based support tool exploiting pre-treatment CT images to predict late xerostomia risk in 3 months after RT in patients with oropharyngeal cancer (OPC). Materials and methods: We performed a multicenter data collection. We enrolled 61 patients referred to three care centers in Apulia, Italy, out of which 22 patients experienced at least mild xerostomia 3 months after the end of the RT cycle. Pre-treatment CT images, clinical and dose features, and alcohol-smoking habits were collected. We proposed a transfer learning approach to extract quantitative imaging features from CT images by means of a pre-trained convolutional neural network (CNN) architecture. An optimal feature subset was then identified to train an SVM classifier. To evaluate the robustness of the proposed model with respect to different manual contouring practices on CTs, we repeated the same image analysis pipeline on "fake" parotid contours. Results: The best performances were achieved by the model exploiting the radiomic features alone. On the independent test, the model reached median AUC, accuracy, sensitivity, and specificity values of 81.17, 83.33, 71.43, and 90.91%, respectively. The model was robust with respect to diverse manual parotid contouring procedures. Conclusion: Radiomic analysis could help to develop a valid support tool for clinicians in planning radiotherapy treatment, by providing a risk score of the toxicity development for each individual patient, thus improving the quality of life of the same patient, without compromising patient care.

A machine learning ensemble approach for 5- and 10-year breast cancer invasive disease event classification.

Designing targeted treatments for breast cancer patients after primary tumor removal is necessary to prevent the occurrence of invasive disease events (IDEs), such as recurrence, metastasis, contralateral and second tumors, over time. However, due to the molecular heterogeneity of this disease, predicting the outcome and efficacy of the adjuvant therapy is challenging. A novel ensemble machine learning classification approach was developed to address the task of producing prognostic predictions of the occurrence of breast cancer IDEs at both 5- and 10-years. The method is based on the concept of voting among multiple models to give a final prediction for each individual patient. Promising results were achieved on a cohort of 529 patients, whose data, related to primary breast cancer, were provided by Istituto Tumori "Giovanni Paolo II" in Bari, Italy. Our proposal greatly improves the performances returned by the baseline original model, i.e., without voting, finally reaching a median AUC value of 77.1% and 76.3% for the IDE prediction at 5-and 10-years, respectively. Finally, the proposed approach allows to promote more intelligible decisions and then a greater acceptability in clinical practice since it returns an explanation of the IDE prediction for each individual patient through the voting procedure.

Pathological Complete Response to Neoadjuvant Chemoimmunotherapy for Early Triple-Negative Breast Cancer: An Updated Meta-Analysis.

Immune checkpoint inhibitors (ICIs) have made a breakthrough in the systemic treatment for metastatic triple-negative breast cancer (TNBC) patients. However, results of phase II and III clinical trials assessing ICIs plus chemotherapy as neoadjuvant treatment were controversial and conflicting. We performed a meta-analysis aimed at assessing the Odds Ratio (OR) of the pathological complete response (pCR) rate in trials assessing neoadjuvant chemoimmunotherapy in TNBC. According to our results, the use of neoadjuvant chemoimmunotherapy was associated with higher pCR (OR 1.95; 95% Confidence Intervals, 1.27-2.99). In addition, we highlighted that this benefit was observed regardless of PD-L1 status since the analysis reported a statistically significant and clinically meaningful benefit in both PD-L1 positive and PD-L1 negative patients. These findings further support the exploration of the role of ICIs plus chemotherapy in early-stage TNBC, given the potentially meaningful clinical impact of these agents. Further studies aimed at more deeply investigating this emerging topic in breast cancer immunotherapy are warranted.

Robustness Evaluation of a Deep Learning Model on Sagittal and Axial Breast DCE-MRIs to Predict Pathological Complete Response to Neoadjuvant Chemotherapy.

To date, some artificial intelligence (AI) methods have exploited Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) to identify finer tumor properties as potential earlier indicators of pathological Complete Response (pCR) in breast cancer patients undergoing neoadjuvant chemotherapy (NAC). However, they work either for sagittal or axial MRI protocols. More flexible AI tools, to be used easily in clinical practice across various institutions in accordance with its own imaging acquisition protocol, are required. Here, we addressed this topic by developing an AI method based on deep learning in giving an early prediction of pCR at various DCE-MRI protocols (axial and sagittal). Sagittal DCE-MRIs refer to 151 patients (42 pCR; 109 non-pCR) from the public I-SPY1 TRIAL database (DB); axial DCE-MRIs are related to 74 patients (22 pCR; 52 non-pCR) from a private DB provided by Istituto Tumori "Giovanni Paolo II" in Bari (Italy). By merging the features extracted from baseline MRIs with some pre-treatment clinical variables, accuracies of 84.4% and 77.3% and AUC values of 80.3% and 78.0% were achieved on the independent tests related to the public DB and the private DB, respectively. Overall, the presented method has shown to be robust regardless of the specific MRI protocol.

An Invasive Disease Event-Free Survival Analysis to Investigate Ki67 Role with Respect to Breast Cancer Patients' Age: A Retrospective Cohort Study.

Characterization of breast cancer into intrinsic molecular profiles has allowed women to live longer, undergoing personalized treatments. With the aim of investigating the relation between different values of ki67 and the predisposition to develop a breast cancer-related IDE at different ages, we enrolled 900 patients with a first diagnosis of invasive breast cancer, and we partitioned the dataset into two sub-samples with respect to an age value equal to 50 years. For each sample, we performed a Kaplan−Meier analysis to compare the IDE-free survival curves obtained with reference to different ki67 values. The analysis on patients under 50 years old resulted in a p-value < 0.001, highlighting how the behaviors of patients characterized by a ki67 ranging from 10% to 20% and greater than 20% were statistically significantly similar. Conversely, patients over 50 years old characterized by a ki67 ranging from 10% to 20% showed an IDE-free survival probability significantly greater than patients with a ki67 greater than 20%, with a p-value of 0.01. Our work shows that the adoption of two different ki67 values, namely, 10% and 20%, might be discriminant in designing personalized treatments for patients under 50 years old and over 50 years old, respectively.

A ultrasound-based radiomic approach to predict the nodal status in clinically negative breast cancer patients.

In breast cancer patients, an accurate detection of the axillary lymph node metastasis status is essential for reducing distant metastasis occurrence probabilities. In case of patients resulted negative at both clinical and instrumental examination, the nodal status is commonly evaluated performing the sentinel lymph-node biopsy, that is a time-consuming and expensive intraoperative procedure for the sentinel lymph-node (SLN) status assessment. The aim of this study was to predict the nodal status of 142 clinically negative breast cancer patients by means of both clinical and radiomic features extracted from primary breast tumor ultrasound images acquired at diagnosis. First, different regions of interest (ROIs) were segmented and a radiomic analysis was performed on each ROI. Then, clinical and radiomic features were evaluated separately developing two different machine learning models based on an SVM classifier. Finally, their predictive power was estimated jointly implementing a soft voting technique. The experimental results showed that the model obtained by combining clinical and radiomic features provided the best performances, achieving an AUC value of 88.6%, an accuracy of 82.1%, a sensitivity of 100% and a specificity of 78.2%. The proposed model represents a promising non-invasive procedure for the SLN status prediction in clinically negative patients.

DeepCEL0 for 2D single-molecule localization in fluorescence microscopy.

MOTIVATION: In fluorescence microscopy, single-molecule localization microscopy (SMLM) techniques aim at localizing with high-precision high-density fluorescent molecules by stochastically activating and imaging small subsets of blinking emitters. Super resolution plays an important role in this field since it allows to go beyond the intrinsic light diffraction limit. RESULTS: In this work, we propose a deep learning-based algorithm for precise molecule localization of high-density frames acquired by SMLM techniques whose ℓ2-based loss function is regularized by non-negative and ℓ0-based constraints. The ℓ0 is relaxed through its continuous exact ℓ0 (CEL0) counterpart. The arising approach, named DeepCEL0, is parameter-free, more flexible, faster and provides more precise molecule localization maps if compared to the other state-of-the-art methods. We validate our approach on both simulated and real fluorescence microscopy data. AVAILABILITY AND IMPLEMENTATION: DeepCEL0 code is freely accessible at https://github.com/sedaboni/DeepCEL0.

Lean Perspectives in an Organizational Change in a Scientific Direction of an Italian Research Institute: Experience of the Cancer Institute of Bari.

Lean management is a relatively new organizational vision transferred from the automotive industry to the healthcare and administrative sector based on analyzing a production process to emphasize value and reduce waste. This approach is particularly interesting in a historical moment of cuts and scarcity of economic resources and could represent a low-cost organizational solution in many production companies. In this work, we analyzed the presentation and the initial management of current ministerial research projects up to the approval by the Scientific Directorate of an Italian research institute. Furthermore, the initial mode in 2021 ("as is") and the potential mode ("to be") according to a Lean model are studied, according to the current barriers highlighted by the final users of the process and carrying out some perspective analyses with some reference indicators.

Early prediction of neoadjuvant chemotherapy response by exploiting a transfer learning approach on breast DCE-MRIs.

The dynamic contrast-enhanced MR imaging plays a crucial role in evaluating the effectiveness of neoadjuvant chemotherapy (NAC) even since its early stage through the prediction of the final pathological complete response (pCR). In this study, we proposed a transfer learning approach to predict if a patient achieved pCR (pCR) or did not (non-pCR) by exploiting, separately or in combination, pre-treatment and early-treatment exams from I-SPY1 TRIAL public database. First, low-level features, i.e., related to local structure of the image, were automatically extracted by a pre-trained convolutional neural network (CNN) overcoming manual feature extraction. Next, an optimal set of most stable features was detected and then used to design an SVM classifier. A first subset of patients, called fine-tuning dataset (30 pCR; 78 non-pCR), was used to perform the optimal choice of features. A second subset not involved in the feature selection process was employed as an independent test (7 pCR; 19 non-pCR) to validate the model. By combining the optimal features extracted from both pre-treatment and early-treatment exams with some clinical features, i.e., ER, PgR, HER2 and molecular subtype, an accuracy of 91.4% and 92.3%, and an AUC value of 0.93 and 0.90, were returned on the fine-tuning dataset and the independent test, respectively. Overall, the low-level CNN features have an important role in the early evaluation of the NAC efficacy by predicting pCR. The proposed model represents a first effort towards the development of a clinical support tool for an early prediction of pCR to NAC.