An electronic portal imaging device (EPID) is an effective detector for in vivo transit dosimetry. In fact, it supplies two-dimensional information, does not require special efforts to be used during patient treatment, and can supply data in real time. In the present paper, a new procedure has been proposed to improve the EPID in vivo dosimetry accuracy by taking into account the patient setup variations. The procedure was applied to the breast tangential irradiation for the reconstruction of the dose at the breast midpoint, Dm. In particular, the patient setup variations were accounted for by comparing EPID images versus digitally reconstructed radiographies. In this manner, EPID transit signals were obtained corresponding to the geometrical projections of the breast midpoint on the EPID for each therapy session. At the end, the ratios R between D(m) and the doses computed by the treatment planning system (TPS) at breast midpoints, D(m,TPS), were determined for 800 therapy sessions of 20 patients. Taking into account the method uncertainty, tolerance levels equal to ± 5% have been determined for the ratio R.The improvement of in vivo dosimetry results obtained (taking into account patient misalignment) has been pointed out comparing the R values obtained with and with-out considering patient setup variations. In particular, when patient misalignments were taken into account, the R values were within ± 5% for 93% of the checks; when patient setup variations were not taken into account, the R values were within ± 5% in 72% of the checks. This last result points out that the transit dosimetry method overestimates the dose discrepancies if patient setup variations are not taken into account for dose reconstruction. In this case, larger tolerance levels have to be adopted as a trade-off between workload and ability to detect errors, with the drawback being that some errors (such as the ones in TPS implementation or in beam calibration) cannot be detected, limiting the in vivo dosimetry efficacy.The paper also reports preliminary results about the possibility of reconstructing a dose profile perpendicular to the beam central axis reaching from the apex to the lung and passing through the middle point of the breast by an algorithm, similar to the one used for dose reconstruction at breast midpoint. In particular, the results have shown an accuracy within ± 3% for the dose profile reconstructed in the breast (excluding the interface regions) and an underestimation of the lung dose.