Phylloids and fibro-adenomas are two benign breast pathologies which had a very close histology at the beginning, yet later their pattern of growth does divert. PT had more positive proliferative cellular nuclear antigen, Ki-67 and androgen receptors than found in FA. Moreover, PT had perforated capsule with finger like projections that is why wide excision surgical procedure is required unlike FA where only simple excision may be needed [10]. Even with the aid of the ultrasound, unless there is a history of rapid increase in the size of the mass; their distinction from each other in many situations does present a dilemma.
In the current work, the performance of the artificial intelligence was studied with regard the diagnosis and discrimination between the PT and FA that presented as masses on mammograms. Ultrasound evaluation was included since it is the descriptive modality which in most situations provides the proper distinction between the variable breast pathologies and sometimes the margins of the masses may be obscured by the glandular tissue on the mammograms.
Previous studies that had focused on the mammograms and AI were concerned mainly with the performance of the AI as a stand-alone screening strategy or as a complementary reading tool to mammogram for the detection of breast cancer in the screening practice not as a tool of disease discrimination [11].
A recent study by Mansour et al. [12] studied the diagnostic performance of the AI- scanned mammograms in correlation with the traditional used conventional breast imaging modalities (the mammogram and the ultrasound) with regard the different breast entities. Another study assessed the impact of the ultrasound artificial intelligence on the differentiation between benign and malignant breast lesions of BI-RADS 4A [13].
Also, there was specific work that considered the potential role of artificial intelligence in the distinction between phyllodes and fibro-adenoma with regard the AI-aided ultrasound [14] or the whole-slide images in core biopsies [15].
The current work was also considered with such differentiation between phyllodes and fibro-adenoma, however it is a leading work to discuss the assessment using the AI- scanned mammograms.
Re-scanning of the mammograms was done through AI in the form of focal color to target the breast mass (already proved as phylloids or fibro-adenoma) on the mammogram that was supported with an auto-applied abnormality scoring percentage of these masses. Then, a correlation between this numerical estimate (i.e., the abnormality scoring percentage) and the pathology results was performed.
Duman et al. [3], suggested that FA could be differentiated from PT by the shape of the tumor, while other groups found no significant difference in the shape between both tumors which tend to be oval or irregular in shape; phyllodes tumors may grow more rapidly than fibroadenomas on follow-up ultrasonography, but they cannot be reliably differentiated by imaging [14, 16, 17].
In this work, PT presented mainly with the rounded /oval shape pattern (61% versus 55% for FA), while the irregular shape went more with the FA (45% versus 39% for PT).
Lee et al. [18] reviewed in accordance with most findings in the literature; that PT presented predominantly with circumscribed margins. This went in concordance with the present work, as circumscribed margins were more common in PT (75.6% versus 61.9% for the FA). However, they disagreed with Duman et al. [3], who stated that circumscribed margins were significantly more common in FA than in PT.
In the study by Wiratkapun et al. [19], 85% of the included PT was complex/heterogeneous masses, as reported in other studies where FA was commonly presented by homogenous texture [3, 19]. The current study also showed that FA was mainly homogeneous (57.2% versus 17.7% for PA).
In this work, when mammograms were scanned with the AI algorithms, correct diagnosis presented in 94.8% (n = 199/210) FA and 89.6% (n = 147/164) PT masses Figs. 1 and 2.
Many benign breast diseases show irregular hypo-echoic masses that can mimic carcinoma on ultrasound [20].
Even when masses displayed irregular shape;
low abnormality scoring percentage- elicited at the AI scanned mammograms- favor benign nature of the mass and so follow up could be recommend to the patient rather than biopsy or unnecessary surgical removal, Fig. 3.
PT commonly is presented as a rapidly growing mass that could be associated with significant painful erythema and warmth of the overlying breast skin [21]. This is a very misleading feature that can delay the diagnosis and sometimes may suggest malignant pathology. However, the diagnosis of PT rather than FA was easily applicable in the current work when the cut off value of the abnormality scoring was more than 49.5%, Figs. 1 and 4.
Such AI related- probability had a positive impact on upgrading the sensitivity and the specificity of the conventional breast imaging from 75.8% and 59.2% to 89.6% and 94.8% respectively.
Large tumor size at presentation or rapid growth raises the suspicion for a phylloids tumor rather than a fibro-adenoma [22]. Fibro-adenoma could show large size as well, which is usually encountered in pregnant or lactating women as their growth is associated with increase in the estrogen, progesterone, and prolactin hormones and at this situation it is termed giant fibro-adenomas. Giant fibro-adenomas could overlap with borderline or malignant phyllodes [23].
In the current study, the accurate diagnosis of large masses more than 5 cm was helped using the abnormality scoring of the AI algorithm into PT or FA, Figs. 1 and 4.
Benign, borderline, and malignant PT of the breast have similar imaging features; some MRI findings can be used to determine the risk of malignancy which include non-circumscribed margins, peri-tumor edema, and low signal intensity on T2-weighted images. On basis of the conventional breast imaging, this task is a challenge [24].
A study performed in 2012, by Dheeba and Selvi [25] showed one of the highest sensitivity (96.9%) and specificity (92.9%) for the proposed AI algorithm in the detection of the cancer in the mammograms. In 2021; the study by Mansour and co-authors [12] showed a sensitivity of 96.8% and a specificity of 90.1% in the discrimination between benign and malignant breast lesions.
Distinction of malignant PA masses that were included in this study was applicable in 100% of these masses (n = 12/12) by the aid of the AI- mammogram combination. This was the condition with masses that presented on mammogram with high density, haziness of the tissue surrounding the tumor due to edema, superimposition of a predominantly intense red color hue on the AI images and a correlating high abnormality scoring suspicion of malignancy that was more than 90%, Fig. 5.
According to Stavros et al. [26], the AI- feature analytic algorithms may support subdividing the BI-RADS category 4, thus can prompt the use of BI-RADS–based structured reporting and encourage the reconsider of tissue sampling for these lesions.
This was the condition with some masses who presented by suspicious features on the primary evaluation by the mammogram and ultrasound, yet when these mammograms were scanned by AI, these masses were not overlaid by color hue and eventually the breast was assigned a low scoring percentage of less than 10% (i.e. lesions of low significance) Figs. 1, 3 and 4. In these cases, if the AI pattern of interpretation were considered in the clinical setting, then biopsy could have been dismissed (especially in case of the small sized masses). Biopsy is warranted if masses presented large size. The interventional procedure in such condition is needed to confirm the absence of malignant potential. The indulgence of the AI findings in the decision of the diagnosis would save the patient from unneeded panic and/or anxiety in large sized benign looking masses.
The heatmap (i.e., color hue) elicited by AI on the scanned mammogram images is used to spot abnormal breast lesions, so it could guide future follow ups or localize lesions that warrant biopsy from those that require further imaging settings. Biopsy is to be considered in case of serial increase in the abnormality scoring percentage of masses under follow up [12].
The current experience showed that masses with low scoring percentage on AI scanned mammograms can then be subjected to interval supervision. Even in case of stationary morphologic features, AI could be considered as a parameter of follow up; in the by monitoring the changes in the intensity of the color hue (i.e., changes from cold light blue or green colors to intense hot colors as yellow, orange, or red) on the mammogram and /or in the value of the abnormality scoring percentage (i.e., increase in the value of the abnormality scoring percentage).
However, precise reports and proper recommendation of the management require continuous supervision of the AI performance by the radiologist [27].
In the current work, AI presented 17 false negative cases that were misdiagnosed as FA instead of PT. That assumption was based on the low abnormality scoring percentage that sometimes presented a value of less than 10%, Fig. 6. The example case: Fig. 6, presented with right breast mass that rapidly increased in size and this was the point that supported the recommendation of the biopsy; otherwise, the mass showed the classic morphologic features of benignity on the baseline as well as the follow up mammograms. Also, the AI scanned images showed a low abnormality scoring that didn’t exceed 25%. That’s why, it is important to keep direct communication with the patient to be accounted with the clinical history and the circumstances of the breast disease.
In breast densities with ACR a and b; masses are obvious with respect to the breast glandular tissue and so mammograms scanned with AI that showed no color hue demarcation and was given a low abnormality scoring of suspicion (i.e., less than 10%) could be considered as a satisfactory modality of scanning. Yet, in case of breast densities assigned ACR c or d patterns; further scanning with ultrasound is required so that not to miss masses overlapped by the dense glandular tissue of the breast and be sure about the BI-RADS category whether normal (category 1) or benign (category 2) Fig. 7.
The current experience showed significant correlation between the BI-RADS category that was assigned by the mammogram and ultrasound combination and the abnormality scoring elicited by the AI scanning with regards the categories; “likely benign, BI-RADS 2” (P value < 0.001), “probably malignant, BI-RADS 4” (P value 0.020) and “likely malignant, BI-RADS 5” (P value < 0.001).
There were limitations to the current work: (1) its retrospective nature limited the ability to determine the Doppler ultrasound details of all the included masses. Resistive index value has not been reported previously and moreover it was not possible to determine whether the distribution vascularity was central or peripheral. (2) Clinical data and demographic findings of the cases were not analyzed, as this study was based only on imaging findings.