The effect of neoadjuvant therapy on radiological, surgical, and pathological results in nonmetastatic breast cancer: A retrospective observational study.

1. Introduction
Breast cancer is the most common malignancy and represents a significant proportion of female cancer morbidity and mortality worldwide.[1] Breast cancer is most commonly seen in the epithelial cells lining the milk ducts (ductal carcinoma), but it can also occur in the breast lobules (lobular carcinoma).[2] Various factors that increase the risk of breast cancer have been clearly identified. In Western countries, thanks to screening programs, a large proportion of breast cancers can be detected at an early stage without symptoms. However, in developing countries, breast cancer usually presents with symptoms such as a lump in the breast or abnormal discharge from the nipple.[3] Breast cancer is diagnosed by physical examination, mammography (MG), or other breast imaging methods and pathological examination of a biopsy sample taken from suspicious tissue.[4]
In the 1980s, neoadjuvant therapy (NAT) was introduced to make locally advanced cancers suitable for surgery and to make breast-conserving surgery (BCS) an option in stages II and III breast cancers.[3,4] Studies have shown that the prognosis is significantly improved in patients who achieve a pathological complete response (pCR) after neoadjuvant chemotherapy or complete disappearance of invasive disease in the breast and lymph nodes.[5,6]
The development of more sophisticated technical possibilities in the field of oncological treatment has led to an increasing number of orders for NAT for patients diagnosed with nonmetastatic breast cancer. NAT usually refers to chemotherapy, targeted therapy, and hormonal therapy.[7,8] NAT is an important part of multimodal therapeutic strategies, allowing tumor shrinkage, reducing the stage of disease, and improving the efficacy of surgery prior to resection.[9,10] NAT is defined according to the biology and molecular subtype of the tumor, according to recommendations for this specific therapeutic approach, such as those proposed by the National Comprehensive Cancer Network.[11]
Radiological imaging plays a vital role in the diagnosis, treatment, and follow-up of breast cancer. Ultrasonography (US), MG, and magnetic resonance imaging (MRI) are the most commonly used imaging modalities to assess tumor size and morphology and treatment response tissue.[11]
However, NAT remains a common practice and is still seen as an active area of research in radiological evaluation. Some studies in the literature have shown that radiological and pathological findings in these patients are not always concordant. As a result, it has been difficult to predict the efficacy of a treatment.[12] This observation can be explained for multiple reasons, including tumor heterogeneity, changes in the biology of residual tumors caused by NAT, and inherent limitations of imaging modalities. Consequently, understanding the impact on radiographic assessment is critical to delivering the best possible treatment strategies and outcomes for patients.
Therefore, a complete review of the NAT effect in terms of radiological assessment is indispensable to improve and optimize the assessment performance of treatment modalities. A large, well-reported study is needed to evaluate the effect of NAT on radiological findings in nonmetastatic breast cancer, its relationship with pathological outcomes, and the results obtained with artificial intelligence (AI). Our data may provide a reference for the acquisition of standardized imaging to assess response and will advocate for selection based on these standards with clinicians, particularly in breast cancer treatment. We will evaluate the performance of imaging-derived response assessments obtained pre- and posttreatment by comparing them with histopathological correlations and AI. This will improve management and imaging outcomes.

2. Material and methods

2.1. Patient selection
This retrospective observational study was designed and included patients between June 2018 and December 2024. In this cross-sectional study, the radiologic, surgical, and pathologic findings of 75 patients with an indication for NAT and 205 patients without an indication for NAT were evaluated. Information about the patients included in the study was obtained from the records of Adiyaman Training and Research Hospital. Informed consent was requested for the procedures conducted during the radiological evaluations; however, it was not required for this observation. As the data were extracted from clinical records, the project was submitted to the Ethics and Research Committee for approval (Approval Date: 18 February 2025 Approval Code: 2025/2-4). Indication for NAT was determined based on tumor size, stage, biological subtype, and patient clinical status using National Comprehensive Cancer Network guidelines.[11] Neoadjuvant therapy included chemotherapy, anti-HER2 therapy, or endocrine therapy protocols, based on the biological characteristics of the tumor, using the same guidelines.

2.1.1. Study inclusion criteria
Patients diagnosed with nonmetastatic breast cancer by percutaneous tru-cut biopsy.

Patients who have had complete radiological evaluations at our hospital.

Patients who have undergone surgery and whose pathological data are available in hospital archives.

2.1.2. Exclusion criteria
Patients diagnosed with DCIS.

Patients diagnosed with metastatic breast cancer.

Patients whose files cannot be accessed from the hospital archives.

Patients whose radiological evaluations are deemed inadequate.

2.2. Radiology
We performed radiological evaluation using a combination of US, MG, and breast MRI. We also included MG in the evaluation of patients diagnosed with breast cancer under the age of 40. In the neoadjuvant treatment group, pretreatment US + MG and breast MRI were blindly evaluated. Post-neoadjuvant treatment radiological evaluation also used imaging techniques to assess the effectiveness of the treatment. These tests provided important information about tumor shrinkage and breast tissue changes, which helped plan subsequent surgical steps. Each radiological evaluation was performed independently and blindly at different times, and the surgical decision was determined for each. Therefore, we performed 2 radiological evaluations in the neoadjuvant treatment group: a pretreatment and a final posttreatment evaluation. Interim radiological evaluations were not included in the study. In the group not receiving neoadjuvant treatment, radiological imaging was evaluated only once with US + MG and breast MRI, and the surgical decision was made based on these blinded evaluations. All radiological evaluations were performed independently and blindly at different times, using the same equipment to avoid bias. Radiological evaluation was performed by M.Ç., a breast radiologist with 14 years of experience in the field. Based on the radiological evaluations, the decision for BCS was made using the American College of Radiology ACR Appropriateness Criteria® guidelines.[13] Images for MG screening were obtained using an IMS GIOTTO TOMO (Bologno, Italy) MG device. A SIGNA Explorer 1.5 Tesla GE (Waukesha) was used for MRI radiological evaluation. All images were interpreted using the institutional database system (Oracle database V1.0.62.409). For US evaluation, patients were evaluated with a LOGIQ GE Ultrasound (USA) device.

2.3. Surgery and pathology
The patients underwent surgical breast-conserving procedures and mastectomies. Patients in the lumpectomy cohort received BCS to excise the substantial primary tumor along with a segment of breast tissue, whereas those having mastectomy underwent subcutaneous, complete, or radical mastectomy procedures. All groups other than BCS were collectively named as Non-BCS. Patients scheduled for BCS were chosen based on the criteria outlined in the review research by Leidenius et al.[14] During the development of BCS, decisions were based on the tumor’s size, location, and spread, as well as whether it would be possible to completely remove the tumor with enough margins and whether it would be a good candidate for radiation, its genetic and molecular characteristics, and the patient’s preferences and concerns about how it would look. Multicentric neoplasms—some situations where BCS should not be used were when the tumor-to-breast ratio was not right, when a positive surgical margin could not be fixed, when radiotherapy was not an option (such as during pregnancy or after previous radiotherapy), when the breast cancer was inflammatory, or when a genetic predisposition meant that both mastoids had to be removed. Surgical procedures were conducted by H.A., a specialist in surgical cancer and general surgery with 25 years of expertise in breast surgery. Specimens from surgically treated patients were assessed by the pathology department, and pathological TN staging was conducted for these patients. The AJCC Cancer Staging Atlas, 9th edition, published by Springer in New York, was utilized for pathological staging.[15] We excluded individuals with incomplete records from the study.

2.4. AI and imaging techniques
AI is significantly transforming the diagnostic imaging environment in healthcare. The AI model was trained using a labeled dataset consisting of imaging data obtained after neoadjuvant treatment and pathological results; during the analysis process, it performed the task of classifying tumor response and comparing treatment efficacy with surgical-pathological findings. This technology integrates machine learning and advanced algorithms. The role of AI in diagnostic imaging is not only about automating processes. AI fundamentally changes the approach to disease diagnosis, making it more efficient and accurate.[16] When traditional image interpretation methods are analyzed, it is determined that these methods are time-consuming and subject to human error. Thanks to AI, images can be processed and analyzed much faster. This significantly reduces the time required to diagnose a patient. This speed is especially important in emergency situations.[17,18]
AI has a key importance in improving patient outcomes, especially in cases such as cancer where early intervention can change the prognosis.[19] AI is effective in the transition to personalized medicine. By analyzing a patient’s specific characteristics and medical history, AI provides specific insights into personalized and effective treatment plans. This personalized approach represents a major advance in healthcare delivery, moving away from a one-size-fits-all model.[20] As a result of the integration of AI into diagnostic imaging, challenges arise. These challenges include potential biases in AI algorithms, concerns about data privacy, and the need for significant investment in training and technology. Clear guidelines and ethical standards are needed for the use and effective management of AI in healthcare.[21] In this study, Python tool was used as an AI model for automated data processing, pattern comparison, and statistical validation.

2.5. Statistical methods
Pearson chi-square test was used to compare categorical data between groups when all expected observation values exceeded 5, but Fisher test was used when any expected observation value fell below 5.[22] We analyzed multiple comparisons using the Bonferroni adjusted Z test and assessed agreement between groups using the kappa coefficient.[23] Changes between combining MG with US and MRI data at baseline and after treatment were analyzed using the McNemar test. Descriptive statistics for categorical variables are reported as frequencies (n) and percentages, while continuous data are presented as mean, standard deviation, and minimum-maximum values. In all analyses and interpretations, the statistical significance threshold was set as P < .05. Statistical analyses were performed using IBM SPSS Statistics 27.0 (IBM Corp. Published in 2020).[24] The data obtained here were analyzed with the Python AI program, and the results obtained were compared. Python-based models were used solely for automated data processing, pattern comparison, fit statistics, and statistical validation.

3. Results
A total of 280 female patients diagnosed with nonmetastatic breast cancer were included in the study, with a mean age of 50.06 ± 11.37 years (range: 21–83 years). Of these, 75 patients (26.8%) received NAT prior to surgery, while 205 patients (73.2%) underwent primary surgery without NAT.

3.1. Overall radiological concordance
Across all patients, a statistically significant correlation was found between the final preoperative MG + US and MRI results (P < .001), with a moderate agreement (κ = 0.607; 95% confidence interval [CI]: 0.54–0.67). In the NAT(+) group, concordance was strong and high (κ = 0.842; 95% CI: 0.78–0.89; P < .001). In the NAT(−) group, the level of agreement decreased to moderate (κ = 0.483; 95% CI: 0.41–0.56; P < .001) (Table 1).

3.2. Breast density and radiological agreement
Agreement between MRI and MG + US varied with breast tissue density:
Type A (predominantly fatty tissue): κ = 0.806; 95% CI: 0.73–0.87; P < .001.
Type B (mostly fatty, less dense): κ = 0.908; 95% CI: 0.85–0.95; P < .001.
Type C (heterogeneously dense): κ = 0.474; 95% CI: 0.38–0.56; P < .001.
Type D (extremely dense): κ = 0.406; 95% CI: 0.33–0.49; P < .001.
This gradient demonstrates that diagnostic concordance declines markedly as breast density increases, indicating that lesion visibility is reduced in dense tissue, particularly on MG + US images (Fig. 1).

3.3. Tumor stage and pathological response
Among patients stratified by tumor stage, significant correlations were found in T2N0, T2N1, and T2N2 subgroups (P < .001 and P = .021). The highest concordance was observed in T2N1 patients (κ = 0.507; 95% CI: 0.42–0.60). Patients achieving pCR after NAT demonstrated strong agreement (κ = 0.704; 95% CI: 0.62–0.78; P = .002) between pre- and postoperative imaging (Table 1).

3.4. Age-related differences
Radiological agreement increased significantly with age: <40 years: κ = 0.497 (95% CI: 0.40–0.59; P = .002), agreement rate = 68.1%, and ≥40 years: κ = 0.612 (95% CI: 0.54–0.68; P < .001), agreement rate = 87.3%.
This confirms that lower breast density in older patients improves inter-modal consistency and diagnostic reliability (Fig. 2).

3.5. Effect of NAT on BCS eligibility
Following NAT, 14.7% (11/75) of patients initially unsuitable for BCS became eligible post-therapy (P = .001; 95% CI: 7.7–24.6%). This demonstrates a statistically significant improvement in BCS feasibility (Tables 2 and 3; Fig. 3).
However, overall BCS rates were 57.1% in non-NAT patients versus 44.0% in NAT-treated patients. The apparent reduction reflects the higher initial tumor stage in NAT candidates rather than reduced treatment efficacy (Fig. 4).

3.6. Radiology-pathology concordance
A weak correlation was observed between final MRI and pathological findings for all patients (κ = 0.045; 95% CI: 0.02–0.07; P < .002).
NAT(+) group: κ = 0.008 (95% CI:–0.01–0.03; P = .012).
NAT(−) group: κ = 0.062 (95% CI: 0.03–0.09; P < .001).
These findings confirm the study hypothesis that “radiologic-pathologic agreement is poor,” emphasizing the limitation of radiology alone in accurately reflecting pathological reality (Table 4). Although the concordance between imaging methods was strong and high in patients undergoing NAT, radiology-pathology concordance remains weak, especially in the post-NAT period.

4. Discussion

4.1. Radiological evaluation and NAT
Among the highlights in our study was that when both mammogram + US and MRI were performed in patients who received NAT, there was greater agreement than in those who did not receive these nanoscale advances. This suggests that NAT-induced tumor shrinkage and morphological changes may facilitate more consistent radiological assessments. These findings align with previous literature emphasizing MRI as the superior modality for evaluating residual disease and planning surgery due to its high soft-tissue contrast.[9,11,25–27] The reduction in TNM stage after NAT likely increases the ability to identify persistent disease, thus resulting in better agreement between imaging modalities.

4.2. Effect of breast density on radiological agreement
We observed that agreement between imaging modalities generally improved with increasing BIRADS density (Types A–D), supporting the established advantage of MRI in dense breast tissue where MG and US are often limited.[12,28–30] Greater difficulties arise in imaging interpretation and establishing a concordance between the various diagnostic models.[28] However, our AI analysis yielded a contradictory result, showing decreased compatibility in the densest breast types (C and D). This discrepancy suggests that while MRI remains the most beneficial tool for dense compositions, high fibroglandular density still poses a significant challenge for standardized interpretation, even for automated algorithms. The results suggest that MRI is expected to be of greater benefit in assessing treatment response in patients with denser breast compositions.

4.3. Age-related variations in radiological concordance
Both clinical and AI analyses demonstrated that radiological concordance significantly increases with patient age, particularly in those over 40. These results seem to be consistent with previous studies showing that younger women tend to have denser breasts, which affects image quality.[7] This may be explained by the fact that older patients often present with less dense breasts, making imaging clearer and more objective for evaluation.[31] In younger patients, fibroglandular composition may vary more between 2 mammograms than in postmenopausal women, who generally have less breast density, allowing lesions to be better visualized on different imaging modalities.[8] Other studies have shown that age-specific changes in breast tissue composition can also affect imaging accuracy.[32–34] Consequently, NAT response assessments may benefit from age-stratified approaches to optimize surgical referrals.

4.4. Impact of NAT on BCS eligibility
A major clinical implication of this study is the role of NAT in converting inoperable cases into candidates for BCS. By reducing tumor burden without compromising oncologic safety, NAT significantly increases BCS rates, a finding supported by both our AI analysis and previous meta-analyses.[6,12,33] Our results reaffirm that NAT is an essential component in achieving favorable surgical outcomes and superior cosmetic results by downsizing tumors that would otherwise require mastectomy.

4.5. Pathological correlation
Although these findings seem positive, the study also addressed the failure to correlate radiological and pathological results. When the analysis performed with AI examined the agreement between radiological findings and pathological results, it was observed that there was a very low correlation in general. As a result of the analysis performed according to AI, it was determined that radiological assessments alone were not sufficient and the need for multimodal imaging techniques was clearly demonstrated. MRI was highly correlated in terms of tumor response and moderate agreement was found for pCRs. These findings are consistent with preclinical data and other studies showing that MRI is the most accurate imaging modality for assessing response but are consistent with the inability of any current scan to accurately predict pCR when considering treatment-induced fibrosis or microscopic residual disease.[9] This reinforces the importance of multimodal assessment, which should integrate imaging and pathological processes to provide better accuracy in defining treatments.

4.6. Clinical implications and future directions
The findings (based on the analysis performed and the AI analysis performed) emphasize multimodal imaging in breast cancer; the approach should be adapted depending on patient characteristics such as age and breast density, whether unimodal or bimodal. The overlap in differential diagnosis and the ease of less invasive surgery can be increased with the use of NAT, but there is still a disconnection between what is seen radiologically and histopathological findings. More research is needed to address these challenges, especially advanced imaging modalities such as radiomics and AI-based algorithms to look at NAT response assessment in more detail, to achieve greater nodal pathological discordance reduction and develop patient-specific surgical strategies.

4.7. Limitations
Although our study, like all studies, has some drawbacks, the study was retrospective and, therefore, selection bias may have affected the sample size, which may affect the results and generalizability of our findings. The fact that all radiological evaluations were performed by a single experienced radiologist, without assessment of intra-observer reliability or inter-observer variability, represents a potential source of bias and is a primary limitation of this study. In addition, the study included the initial and final radiological assessment of all patients. Intermediate radiological assessments were not mentioned in our study. These limitations may be addressed in the future with additional prospective large-volume studies that include standard imaging protocols and advanced computational techniques. Our study did not include molecular subtyping of breast cancer. The molecular subtype of a tumor plays an important role in the response to NAT and affects imaging concordance and histopathological parameters. Future studies examining the relationship between molecular subtypes and imaging concordance should allow for improved response assessment.

5. Conclusion
Our study highlights the essential role of NAT in nonmetastatic breast cancer, optimizing both radiological assessment and surgical planning through the use of AI. The improved level of agreement between imaging modalities highlights the need for multimodal imaging in post-NAT evaluation. This study adds to the growing evidence for the role of NAT in improving imaging accuracy, increasing breast conservation fidelity, and facilitating surgical planning. However, significant differences remain between radiographic and histopathological assessments, necessitating the need for complementary assessment approaches. Future directions include further developments in breast cancer treatment strategies resulting from increased molecular profiling and the emerging use of new imaging technologies and immune markers to better select patients suitable for NAT.
The quality, reliability, and quantity of data are critical factors that directly affect the success of AI methods. In today’s world where the importance of genetic information is increasingly understood, radiomic models that combine genomic profiles, histological data, biomarkers, and patient history have the potential to provide much more accurate and reliable results. In recent years, we have witnessed the evolution of machine learning models toward deep learning models. With the increase in computer processor speeds and especially the introduction of graphic processing units, studies on artificial neural networks have accelerated and various innovative models have been developed in this field. However, in order for these new technologies to effectively support clinicians in the field of radiology, interdisciplinary collaboration and cooperation are of great importance.

Acknowledgments
We would like to thank the Medical Oncology Department and Pathology Department of our hospital for their contributions to our study. We would also like to thank the developers of the Python program for their contribution to our work.

Author contributions
Conceptualization: Mahmut Çorapli, Hüseyin Alakuş.
Data curation: Mahmut Çorapli.
Formal analysis: Mahmut Çorapli.
Investigation: Mahmut Çorapli.
Methodology: Mahmut Çorapli.
Project administration: Mahmut Çorapli, Hüseyin Alakuş.
Software: Mahmut Çorapli.
Supervision: Mahmut Çorapli, Hüseyin Alakuş.
Validation: Hüseyin Alakuş.
Visualization: Hüseyin Alakuş.
Writing – original draft: Mahmut Çorapli.
Writing – review & editing: Mahmut Çorapli.