Breast sarcomas are among the rarest and most aggressive malignancies encountered in breast imaging. They account for less than 5% of all soft tissue sarcomas and less than 1% of all malignant breast neoplasms. Unlike the common breast carcinomas that radiologists routinely evaluate, sarcomas arise from the mesenchymal tissue of the breast rather than the epithelium, giving them fundamentally different biological behaviors and imaging appearances. Common histopathological subtypes include fibrosarcoma, angiosarcoma, and pleomorphic sarcoma, with angiosarcoma being the most frequent. These tumors enlarge rapidly, metastasize hematogenously rather than via lymphatics, and carry a 10-year mortality rate of approximately 50%.
Primary vs. secondary sarcoma: The clinical literature distinguishes between primary breast sarcoma (PBS), which arises de novo in the breast without a preceding insult, and secondary breast sarcoma (SBS), which typically develops as a late complication of prior radiation therapy to the breast. The two subtypes carry starkly different prognoses: PBS has an estimated median survival of 107 months, while SBS has an average of only 45 months. Both subtypes share an estimated annual incidence of approximately 4.6 cases per million women, confirming that any single institution will accumulate very few cases over many years of practice.
The imaging problem: Breast sarcomas present with non-specific clinical and radiological features that frequently overlap with benign and malignant lesions more commonly encountered in breast imaging. The absence of a palpable lump in a meaningful proportion of cases, combined with the potential for skin-only presentations, creates a diagnostic environment in which sarcomas are easily missed or misclassified. Prior studies have reported discrepant imaging descriptions across institutions, likely because no single center accumulates enough cases to establish robust imaging criteria.
This 2025 retrospective imaging case series from the Icahn School of Medicine at Mount Sinai, published in Cancers, addresses this gap by systematically characterizing the clinical presentation, histopathology, and multi-modality imaging features of 18 biopsy-proven breast sarcomas accrued over a 16-year period (2008 to 2024). It is also the first published study to evaluate the performance of an ultrasound AI decision support system specifically on breast sarcoma lesions.
The study used an IRB-approved retrospective design. The investigators queried a radiology database from 2008 to 2024 using the search term "breast sarcoma," applying consecutive sampling to mitigate selection bias. The inclusion criterion required histologically confirmed breast sarcoma with imaging available prior to surgical resection. Of 22 patients identified, 4 were excluded due to lack of pre-resection imaging, yielding a final cohort of 18 patients. Informed consent was waived given the retrospective design. Exclusion criteria were deliberately narrow to maximize capture of imaging variability across the full clinical spectrum of the disease.
Imaging modalities and equipment: Mammography was available for 13 lesions, performed on Hologic Selenia units at the institution, with 5 external mammograms submitted for internal review. Ultrasound was available for 19 lesions (including 2 recurrent cases), using Philips IU22 and Epiq Elite systems with linear transducers at 10 to 18 MHz, capturing grey-scale and color Doppler images in radial and anti-radial views. Breast MRI with contrast was available for 9 lesions, acquired on 3T GE Discovery MR750, 1.5T GE Signa, or 1.5T Siemens Magnetom Sola scanners, using dynamic contrast-enhanced fat-saturated axial imaging sequences.
Imaging analysis: All images were reviewed retrospectively by two radiologists (DW and CC), each with 8 years of experience, using the Breast Imaging Reporting and Data System (BI-RADS) 5th Edition lexicon. Readings were performed independently prior to consensus, and 100% inter-reader agreement was achieved after discussion. Lesion size was based on the maximum diameter from ultrasound measurements, selected because ultrasound data were available for all cases. For patients presenting with only skin thickening and no discrete intra-mammary lesion, no lesion size was recorded.
AI analysis protocol: KOIOS AI Decision Support (DS), version 3.6.1 (KOIOS Medical Inc., New York), was applied retrospectively to all available ultrasound images, including those from outside facilities. A single radiologist (CC) manually placed a region of interest (ROI) over the biopsy-proven lesion in orthogonal views on pre-biopsy B-mode greyscale images. The KOIOS system generated a probability of malignancy stratified into four BI-RADS-aligned categories: probably benign (3), low suspicion (4A), moderate suspicion (4B to 4C), and highly suspicious (5). Statistical analysis was limited to descriptive statistics, with continuous variables expressed as mean plus or minus standard deviation and categorical variables as percentages with 95% confidence intervals.
The cohort comprised 17 female patients with a mean age of 53.9 years (SD 20.8, range 20 to 80 years) and one male patient aged 72 years. Nine patients had PBS and nine had SBS. Disease recurrence occurred in 2 of 18 patients (11.1%), one in each group. The wide age range, from 20 to 80 years, underscores that breast sarcoma does not follow the age-dependent incidence pattern of common breast carcinomas, and the inclusion of a pregnant patient (22-year-old with high-grade angiosarcoma) and a male patient illustrates the breadth of affected demographics.
Clinical presentation: Of the 18 patients, 9 (50%) presented with a palpable lump, 5 (27.8%) with breast skin changes alone, 2 (11.1%) with both a palpable lump and overlying skin changes, 1 (5.6%) with asymmetric breast enlargement, and 1 (5.6%) with increased nipple sensitivity. Three patients had no identifiable intra-mammary finding on imaging, presenting only with skin thickening. All three of these patients had a prior history of breast radiation therapy and were ultimately diagnosed with secondary angiosarcoma, representing a clinically important pattern that radiologists must recognize to avoid delayed diagnosis.
Histopathological subtypes: Angiosarcoma was by far the most common subtype, accounting for 13 of 18 cases (72.2%). This was followed by spindle cell sarcoma in 2 cases (11.1%), and one case each of granulocytic sarcoma, high-grade pleomorphic sarcoma, and fibromyxoid sarcoma (5.6% each). The predominance of angiosarcoma is consistent with broader literature on breast sarcomas and partly reflects the inclusion of radiation-associated SBS cases, where angiosarcoma is the typical histological outcome of prior radiation-induced vascular injury.
Immunohistochemistry (IHC) profiles varied by subtype. The granulocytic sarcoma case showed CD68+, CD34+, MPO+, lysozyme+, CD43+, and CD117+ staining, consistent with myeloid origin. The pleomorphic sarcoma in the male patient showed desmin positivity with scattered ER+ and PR+ stromal cells. Average lesion size across the cohort was 2.6 cm by imaging (range 0.5 to 7.1 cm, 95% CI 1.69 to 3.51 cm), reflecting a wide spectrum of tumor burden at diagnosis.
Thirteen cases underwent mammographic evaluation. The distribution of mammographic findings was notably more variable than what has been reported in prior smaller series. Three of 13 cases (23.1%) presented as an irregular, indistinct, equal- or high-density mass, the appearance most classically associated with malignancy in standard breast imaging. Another 3 of 13 (23.1%) presented as focal or global asymmetry without calcifications, a far subtler and potentially easily overlooked finding. One case (7.7%) presented as architectural distortion, and one (7.7%) showed asymmetry with amorphous calcifications.
Negative and non-specific findings: Two cases (15.4%) had negative mammograms, meaning no abnormality was detectable despite the presence of histologically confirmed sarcoma. Three additional cases (23.1%) showed only skin thickening without an identifiable intra-mammary mass. Taken together, these data mean that mammography was unable to demonstrate a discrete mass in approximately 38% of cases in this cohort, a striking miss rate that reflects both the imaging characteristics of sarcoma and the known limitations of mammography for certain lesion types.
Comparison with prior literature: Most prior descriptions of breast sarcoma on mammography emphasize well-circumscribed, oval-shaped masses. The distribution observed here is notably broader, with irregular masses, asymmetries, and architectural distortion all represented. The authors attribute this variability to the longer duration and larger sample size of their series compared to earlier reports, suggesting that older descriptions based on very small case series may have captured only a subset of the true imaging spectrum.
The clinical implication is that radiologists should not rely on well-circumscribed morphology as the expected mammographic appearance of sarcoma. Focal asymmetry and architectural distortion, especially in patients with a history of prior breast radiation, should prompt consideration of sarcoma in the differential diagnosis regardless of mammographic appearance.
Ultrasound was available for 19 cases, including 2 recurrent lesions. A discrete mass was identified in 16 of 19 cases (84.2%). One case (5.2%) showed only skin thickening without a discrete mass, and no suspicious sonographic finding was identified in 2 cases (10.5%). Among the 16 cases with a mass, shape was irregular in 13 (81.2%) and oval in 3 (18.8%). Masses were predominantly parallel in orientation and hypoechoic in echogenicity.
Margin and posterior features: Non-circumscribed margins were the dominant pattern, encompassing obscured, indistinct, and irregular subtypes. Posterior acoustic enhancement was a notable and somewhat unexpected feature, present in multiple cases. This finding is more commonly associated with cystic or vascular lesions than with solid malignant masses, and its presence contributed to at least one diagnostic challenge in the cohort. Internal vascularity on color Doppler was identified in several cases, a feature that has been under-discussed in prior breast sarcoma literature.
The cyst mimic: One particularly instructive case involved a 29-year-old woman with a 5.5 cm oval, circumscribed, parallel hypoechoic mass with mixed solid and cystic components and internal vascularity. The ultrasound appearance raised the possibility of a complex cystic lesion, a finding that would typically not prompt immediate biopsy concern in a young woman. This case was the only one misclassified by the AI system, with KOIOS DS categorizing it as probably benign (BI-RADS 3) rather than suspicious. Subsequent contrast-enhanced MRI showed an irregular mass with heterogeneous enhancement, and biopsy yielded high-grade sarcoma. This case illustrates both the challenges posed by atypical sarcoma morphology and the limitations of AI systems trained predominantly on common carcinoma subtypes.
The ultrasound findings in this cohort, particularly the predominance of irregular shape, non-circumscribed margins, and posterior acoustic enhancement, contrast with prior series that described breast sarcomas as more commonly oval or round with circumscribed margins. As with the mammographic findings, the authors attribute this discrepancy to the broader and more representative sampling achieved through their longer accrual period.
Contrast-enhanced breast MRI was available for 9 cases. Six of 9 (66.7%) presented as an intramammary mass, most commonly displaying irregular shape, irregular margins, and heterogeneous enhancement. These features are consistent with the malignant kinetics and morphological criteria used in routine breast MRI interpretation and align with the BI-RADS 4 to 5 category descriptors. Three of 9 cases (33.3%) presented as skin thickening and enhancement without any associated intramammary mass on MRI, again reflecting the skin-dominant presentation characteristic of radiation-induced secondary angiosarcoma.
Comparison with prior MRI series: The shape and margin findings in this cohort are concordant with the Wienbeck et al. multicenter study, which reported irregular morphology as the dominant MRI pattern for breast sarcomas. The findings differ from Smith et al., where round or oval mass morphology was more commonly described. The heterogeneous enhancement pattern observed across most masses in this series reflects the cellular and vascular heterogeneity inherent to high-grade sarcomas, distinguishing them from the more homogeneous enhancement sometimes seen in lower-grade lesions.
Clinical role of MRI: Although MRI was not available for all patients (reflecting real-world clinical practice rather than a standardized protocol), it provided the most detailed morphological and enhancement characterization of any modality in this series. For cases where mammography and ultrasound provided equivocal or negative results, MRI offered the most complete depiction of tumor extent and tissue involvement. The three skin-only cases on MRI illustrate that even MRI cannot always identify a discrete intra-mammary mass in radiation-associated angiosarcoma, where the tumor may be diffuse and primarily dermal rather than parenchymal.
From a practical standpoint, the high rate of heterogeneous enhancement and irregular morphology on MRI supports the use of dynamic contrast-enhanced MRI as a staging and characterization tool for suspected breast sarcoma, particularly when other modalities are equivocal or when the lesion's full extent needs to be defined before surgical planning.
KOIOS AI DS assessment was available for 16 ultrasound cases with biopsy-proven breast sarcoma. This is the first published analysis of any ultrasound AI system applied specifically to breast sarcomas. In 14 of 16 cases (87.5%, 95% CI 0.62 to 0.98), the AI classified the lesion as suspicious (BI-RADS 4A to 4C). In 1 of 16 cases (6.3%, 95% CI 0.002 to 0.30), the AI classified the lesion as probably malignant (BI-RADS 5). Combined, the AI correctly flagged the lesion as requiring tissue sampling in 93.8% of cases.
The misclassification: One case (6.3%) was classified by KOIOS as probably benign (BI-RADS 3). This was the 29-year-old woman whose sarcoma presented as a large, oval, circumscribed, mixed solid-cystic mass with posterior acoustic enhancement. The AI system classified this as probably benign, generating a misclassification that could have resulted in a 6-month follow-up interval rather than immediate biopsy. The authors note that this atypical presentation closely mimicked benign complex cystic lesions, a category well-represented in KOIOS training data. The subsequent MRI finding of heterogeneous enhancement prompted biopsy, which confirmed high-grade sarcoma.
Comparison with carcinoma performance: AI DS systems like KOIOS have reported sensitivities of 0.95 to 0.98 for breast carcinomas in prior studies. The 93.8% sensitivity observed here for sarcomas is nominally lower but broadly comparable, particularly given the small sample size and the inclusion of a morphologically atypical case that would challenge human radiologists as well. The KOIOS algorithm was trained on over 450,000 breast ultrasound images from more than 30 clinical sites covering 40 ultrasound machine models. However, the sarcoma-specific component of that training set is not publicly disclosed, making it impossible to determine whether sarcomas are adequately represented.
Limitations of AI on skin-thickening-only cases: For cases where breast sarcoma presented solely as skin thickening without a discrete intra-mammary mass, the KOIOS system was unable to generate a result because ROI placement requires an identifiable lesion. This is an important practical limitation: AI decision support systems designed for focal lesion analysis are structurally unable to evaluate diffuse or skin-predominant presentations, which represent a diagnostically challenging and clinically significant subset of breast sarcoma.
Retrospective design and sample size: The retrospective nature of the study and the small sample of 18 patients represent the primary methodological constraints. While this cohort is described as one of the largest recent breast sarcoma imaging series given the rarity of the disease, comparative statistical analysis between PBS and SBS subtypes is underpowered. Descriptive statistics without inferential analysis are appropriate given these constraints, but they limit the generalizability of any apparent subgroup differences observed in the data.
Equipment and institutional variability: Ten of 19 ultrasound examinations and 5 of 13 mammograms were performed at outside institutions, introducing variability in acquisition parameters, machine models, and image quality. The authors acknowledge that this "decreased standardization of technique and image acquisition" is a limitation, though they argue it also reflects real-world clinical practice where imaging from referring institutions is routinely submitted for review. The non-blinded nature of the radiological review, where reviewers were aware of the sarcoma diagnosis, introduces the potential for confirmation bias in feature attribution.
AI analysis caveats: The retrospective ROI placement by a single radiologist introduces selection bias, as the radiologist knew the lesion location and pathology result at the time of ROI placement. The authors cite prior work by Barinov et al. suggesting that ROI boundary variation has minimal impact on KOIOS category assignment, supporting the validity of the approach. However, the prospective performance of KOIOS on breast sarcomas, where the radiologist would not know the diagnosis in advance, remains untested.
Clinical implications and future work: Despite these limitations, the study provides actionable guidance for clinical practice. Radiologists evaluating breast lesions should recognize that sarcoma can present across a wide range of mammographic and sonographic morphologies, that skin-thickening-only presentation in post-radiation patients warrants high suspicion for angiosarcoma, and that AI decision support performs comparably to its carcinoma benchmarks for most but not all sarcoma presentations. Future prospective studies incorporating larger multicenter cohorts, standardized imaging protocols, and AI systems trained with transparent sarcoma-specific data representation are needed to establish more definitive performance benchmarks.