Quantitative ultrasound imaging for predicting response and guiding personalized neoadjuvant chemotherapy in breast cancer: randomized phase 2 clinical trial results

Abstract

Quantitative ultrasound (QUS) detects early tumor microstructural changes during neoadjuvant chemotherapy (NAC), enabling personalized treatment adaptation. This study assessed the accuracy of machine learning models using serial QUS data to predict treatment response and evaluated their feasibility for guiding treatment personalization. This single-center, phase 2 randomized controlled trial (clinicaltrials.gov NCT04050228, Dec/2019) enrolled stage II-III breast cancer patients planned for standard NAC. QUS imaging was performed at baseline and week 4, with the latter used for response prediction. Patients were randomized 1:1 to standard or experimental arms, stratified by hormone receptor status. In the standard arm, oncologists were blinded to QUS results. In the experimental arm, predictions were disclosed to allow treatment modification at week 4. Final response was determined histopathologically (>30% tumor reduction or <5% cellularity). Between June 2018 and September 2023, 146 patients were enrolled, and 120 randomized (standard: 57, experimental: 63). Response rates were 93.0% (standard) and 96.8% (experimental). The model achieved 92% accuracy, 83% sensitivity, 93% specificity, and 99% positive predictive value. In the experimental arm, 8/63 patients were predicted non-responders, with 4 undergoing treatment modification. QUS-based machine learning enables accurate early response prediction and supports adaptive treatment strategies in future trials.

Fig. 1

Trial consort diagram.

Fig. 2. B-mode imaging and corresponding QUS-parametric maps (colour coded) at different experimental times (pre-treatment or week 0 (W0), and week 4 of neoadjuvant chemotherapy (w4)) for one representative patient from responder (R) and non-responder (NR) groups.

The scale bar represents a length of 2 cm. The colour bar indicates the range of parametric colour overlay: MBF range was from −9.6 dB to 34.0 dB, SS range was from −5.7 dB/MHz to 1.6 dB/MHz, SI range from −7.3 dB to 49.0 dB, AAC range was from 20.2 db/cm³ to 81.6 db/cm³, ASD range was from 40 to 171 μm. MBF mid-band fit, SS spectral slope, SI spectral intercept, AAC average acoustic concentration, ASD average spectral diameter.

Fig. 3. Detailed class score plot.

Individual patient predictions based on predictor class scores at week 4 for patient response. R indicates the zone (negative class score) for predicted response, and NR indicates the zone (+ve class score) for non-response. Responder, complete responder and non-responder are defined as per RECIST criteria. Weak responders were borderline in regards to size changes (±5%) or had significantly diminished cellularity (5% or less) but did not meet size criteria.

Fig. 4. Trial Sankey diagram.

Sankey diagram for predicted response at week 4 using QUS-radiomics model with the final response on an individual patient basis. Individual tiles indicate response type (actual (A) or predicted (P)) and tumour histopathological type. In the experimental arm, 59/63 patients were not adapted based on information provided to their oncologist, whereas 4/63 were adapted. Colour tiles in the first two columns indicate complete responders, responders, and non-responders as per RECIST. Weak responders were borderline in regards to size changes (±5%) or had significantly diminished cellularity (5% or less) but did not meet size criteria. The third column indicates tumour histopathological type (invasive ductal carcinoma, mucinous type, invasive lobular carcinoma, ductal carcinoma in situ in over 50% of tumour volume, micropapillary type and other).

Fig. 5. Clinical histories of adapted patients.

A Schematic diagram for the administration of standard AC-T chemotherapy. Weeks are shown from left to right. Typical durations are illustrated. B–E The four patients in the Experimental Arm that were adapted (EAA-26, EAA-27, EAA-28, and EAA-57) in Fig. 4. In (B), AC was shortened to move to T; in (C), the AC and T were shortened to move to surgery; in (D), AC was shortened and T was intensified; and in (E), FEC-D was stopped early due to prediction and full clinical response.

Fig. 6. ROI size change, (W4 – W0)/W0, versus W4 Prediction Class Score.

Size changes at Week 4 versus Ultrasound Class Score. This plot presents tumour size change between week 4 and week 0 per tumour versus predictive class score for response. Responder, complete responder and non-responder are defined as per RECIST criteria. Weak responders were borderline in regards to size changes (±5%) or had significantly diminished cellularity (5% or less) but did not meet size criteria.