The main purpose of this study is to develop a computer-aided prediction model for NAC treatment response. Based on the heterogeneity of internal parametric tumor composition commonly observed, this study will utilize the histologic characteristics and treatment response to investigate the image features as input data for predicting treatment response using Deep Learning technology. Using this technique, preoperative treatment evaluation may be facilitated by tumor heterogeneity analysis from developed dynamic radiomics, and the possibility of personal medicine can be realized not far ahead. In the first two years of this study using images from DCE-MRI, PET/CT and QDS-IR, we plan to develop the image processing algorithms, including segmenting breast and tumor region, extracting image feature which reflects angiogenic properties and permeability of tumor, which are highly correlated with NAC treatment response. During the third year of the project, the morphology and texture features from first two years can be combined for PET/MRI and prediction model can be achieved in accordance with the features extracted from dynamic features extraction using longitudinal images of PET/MRI.

Breast cancer is the most frequently diagnosed cancer and remains the fourth leading cause of cancer deaths in Taiwan women over the past decade. Decisions about the best treatment for breast cancer is based on the result of estrogen (ER) and progesterone receptor (PR) test, human epidermal growth factor type 2 receptor (HER2) test, and TNM staging using biopsy. After evaluation of menopause status and response of ER, PR and HER2, the treatments for stage 2 or above breast cancer may consider neoadjuvant chemotherapy (NAC) for the benefits of (1) converting an inoperable to a surgical resectable cancer, (2) metastasis management, (3) shrink the tumor, (4) improved overall survival and recurrence free survival rate (5) histologic parameters predictive. It is known that patients with pathological complete response (pCR) after NAC are associated with better disease-free survival and improved overall survival. Therefore, it is essential to develop more effective regimens and stratify patients based on computer assisted prediction model to evaluate the response of NAC.

The main purpose of this study is to develop a computer-aided prediction model for NAC treatment response. Based on the heterogeneity of internal parametric tumor composition commonly observed, this study will utilize the histologic characteristics and treatment response to investigate the image features as input data for predicting treatment response using Deep Learning technology. Using this technique, preoperative treatment evaluation may be facilitated by tumor heterogeneity analysis from developed dynamic radiomics, and the possibility of personal medicine can be realized not far ahead. In the first two years of this study using images from DCE-MRI, PET/CT and QDS-IR, we plan to develop the image processing algorithms, including segmenting breast and tumor region, extracting image feature which reflects angiogenic properties and permeability of tumor, which are highly correlated with NAC treatment response. During the third year of the project, the morphology and texture features from first two years can be combined for PET/MRI and prediction model can be achieved in accordance with the features extracted from dynamic features extraction using longitudinal images of PET/MRI. The followings are the expected contributions:

To propose a novel parametric dynamics features for overcoming the issues with traditional thresholding method.

To develop segmentation algorithms for breast tissue and tumor region on DCE MRI in order to improve treatment response prediction.

To develop trajectory analysis for non-invasive QDS-IR image To develop segmentation algorithm for metabolic tumor volume by registering PET uptake boundary with CT tumor boundary in order to improve reliability and reproducibility of morphology feature.

By combining preceding techniques developed for PET/CT and DCE MRI, new types of features, namely the parametric dynamics features from PET/MRI can be served as dynamic radiomics for predicting NAC response.

To develop a Deep-Learning algorithm, which is the essential of the project in terms of self-learning methodology, for exploiting the high dimensional features space in search of the prediction model.

Subject Eligibility and Enrollment This prospective study was approved by the local Institutional Ethics Committee who waived the requirement for patient approval and written informed consent for the review of records and images.

From April 2015 to June 2019, 60 women with breast cancer who underwent NAC were screened for eligibility. The eligibility criteria were patients who: (a) were > 18 years of age, (b) with pathologically confirmed breast cancer with core needle biopsy, (c) were willing to undergo NAC, (d) were eligible for surgery after NAC, and (e) were willing to undergo at least two PET/MR scans during NAC: the first [R0], pre-treatment; and the second [R1], after two cycles of chemotherapy (post-treatment). Of the 60 women, 14 were excluded for the following reasons: (a) distant metastases found after the first scan (n=1), so that they were not eligible for surgery; (b) patients unable to complete two sequential PET/MR scans for personal reasons (n=10); (c) patients refused further NAC after the first cycle (n=2); and (d) post-treatment scan could not be performed because of elevated creatinine level (n=1). The study inclusion flow diagram is shown in Figure 1. All patients received 6 to 8 cycles of NAC including the following options. For human epidermal growth factor receptor 2 (HER2)-negative patients, NAC included the concomitant use of epirubicin with cyclophosphamide (EC based) or epirubicin with cyclophosphamide followed by taxanes (docetaxel or paclitaxel; EC+T). For HER2-positive patients, the NAC protocols included the concomitant use of taxanes with trastuzumab (TH based), or a concurrent regimen of taxanes and cyclophosphamide followed by trastuzumab (EC+TH), or taxanes with trastuzumab and pertuzumab (THP). One triple-negative patient received paclitaxel with afatinib.

PET/MR image acquisition MR examinations were performed on a 3T PET/MR scanner (Biograph mMR, Siemens Healthcare, Erlangen, Germany). The midtreatment scan was arranged after two cycles of NAC. The last scan was arranged before surgery. The patients were scanned in the prone position using a dedicated four-channel breast coil (Noras GmbH, Höchberg, Germany). After fasting for at least 6 hours, the patients were injected intravenously with 3.7-5.55 MBq/kg (0.1-0.15 mCi/kg) of 18F-fluorodeoxyglucose (18F-FDG). PET/MR scanning was performed approximately 60 minutes later. A 10-minute breast PET was performed along with precontrast breast MRI in prone position. The PET images were reconstructed with an ordered-subset, expectation-maximization iterative algorithm (3 iterations, 21 subsets), with a 4mm post reconstruction Gaussian filter and an image matrix of 172 x 172. Attenuation correction of PET data was obtained by using a 4-tissue-class filter (air, lung, fat, soft tissue) segmented attenuation correction map , which was reconstructed from a 2-point Dixon MR pulse sequence.

The preconstrat breast MRI protocol included 2D gradient echo T1-weighted (repetition time [TR]/echo time[TE],500/9.8 ms, flip angle 150°, slice thickness 3.5 mm; matrix field of view [FOV] 320 × 320), T2-weighted short tau inversion recovery (STIR) images (TR/TE/TI = 3000/78/230 ms, matrix size 320 × 320, slice thickness 4.0 mm, matrix FOV 330 × 330 mm), diffusion-weighted images (DWIs) (TR/TE = 7500/83 ms, matrix size 192 × 77, slice thickness 4.0 mm, FOV 360 × 1830 mm, b value = 50, 600 and 1000 sec/mm2, average = 2) with apparent diffusion coefficient (ADC) maps were acquired.

Seven dynamic contrast-enhanced images, including one before and six after contrast agent administration, were acquired using a fat-saturated T1-weighted fast low angle shot (FLASH) 3D gradient echo sequence (TR/TE 4.1/1.5 ms, slice thickness and gap 1.0/0 mm, matrix size 352 × 282, flip angle 10°, 146 slices, FOV 330 × 330 mm) with spectral attenuated inversion recovery (SPAIR), each set took approximately 60 seconds. The postcontrast DCE-MRI was started 10 seconds after intravenous injection of 0.1mmol/kg gadobutrol (Gadovist; Bayer Pharma AG, Berlin, Germany) with an injection rate of 3 mL/s.

Histopathology assessment The histologic type and grade were based on histopathological reports of ultrasound-guided core biopsies performed before NAC. The expressions of estrogen receptor (ER), progesterone receptor (PR), and HER2 were assessed using immunohistochemical staining. ER and PR positivity were assessed using the Allred score, in which a score > 3 was considered to be positive. Tumors were considered to be HER2 positive if the immunohistochemical score was 3+. In cases with an equivocal HER2 status (score 2+) in immunohistochemistry, fluorescence in situ hybridization analysis (FISH) was performed to confirm the diagnosis.

A pCR was defined as the absence of residual invasive cancer except for ductal carcinoma in situ (DCIS) in the surgical breast specimen after NAC and the absence of axillary lymph node involvement, as previous reported by Pinder et al.

Image analysis Images were interpreted by two radiologists (Y.F.L. and N.C., both with 12 years of experience in breast imaging). Tumor appearance on MRI (mass, non-mass enhancement) was visually assessed by the two radiologists. The regions of interest (ROIs) were manually defined by selecting a two-dimensional region enclosing the largest cross-sectional area of high signal intensity of enhancement that differed from normal background parenchyma on the fifth contrast-enhanced series, while avoiding normal parenchyma and marking clip artifacts . If the cancer was multifocal or multicentric, the ROI was measured at the tumor area with the largest size.

After the tumor ROIs had been segmented, PET and MR images including DWI, and apparent diffusion coefficient (ADC) image ROI measurements were aligned using the ROIs drawn on dynamic contrast enhanced (DCE) images using semi-automated non-rigid registration and manual alignment where necessary. The ROIs were simultaneously evaluated on six different phases of DCE images (coded as P1 to P6). The ROIs were also registered onto other series including DWIs at b value=1000 sec/mm2, ADC map and PET images . From each ROI, six first-order and four second-order texture features were automatically computed. Histogram analysis was used to assess first-order textural features including mean, standard deviation (SD), median, 5th and 95th percentile values, kurtosis, and skewness. For second-order textural analysis, a gray-level co-occurrence matrix (GLCM) was used as the parent matrix, and a four heterogeneous textural features were extracted from the GLCM, namely difference entropy (DiffEntropy), difference variance (DiffVariance), contrast, and entropy [18]. These textural parameters were calculated using MR Multiparametric Analysis prototype software (Siemens Healthcare, Erlangen, Germany). All imaging parameters were evaluated using PET/MR scans performed pretreatment (R0) and midtreatment (R1), and before surgery (R2) are analyzed.

Inclusion Criteria:

• (a) were > 20 years of age,
• (b) with pathologically confirmed breast cancer with core needle biopsy
• (c) were willing to undergo NAC
• (d) were eligible for surgery after NAC
• (e) were willing to undergo at least three PET/MR scans during NAC: the first [R0], pre-treatment; and the second [R1], after two cycles of chemotherapy (post-treatment) and before surgery [R2]

Exclusion Criteria:

• (a) distant metastases or recurrent breast cancer.
• (b) unable to comply with sequential PET/MR scanning schedule.
• (c) Impaired renal function, CCR>30ml/min.
• (d) Known aller