Matthew Olmstead – mjo5585@psu.edu

Instagram: @mattomatty707
Graduate Program in Acoustics, The Pennsylvania State University, University Park, PA, 16802, United States

Hyungjoo Park – hpp5133@psu.edu
MD Rizwanul Kabir – rizwanulkabir@vt.edu
Aiguo Han – aiguohan@vt.edu
Yun Jing – yqj5201@psu.edu

Popular version of 1aBAa7 – Improving Photoacoustic Imaging through the Skull using Deep Learning: Considering 3D Effects
Presented at the 190th ASA Meeting
Read the abstract at https://eppro01.ativ.me/web/page.php?page=Session&project=ASASPRING2026&id=4082496&nohistory&nohistory=true

–The research described in this Acoustics Lay Language Paper may not have yet been peer reviewed–

Did you know that there’s potentially a better way than ultrasound to image your brain? Say hello to photoacoustics, which combines the benefits of optics and sound, giving us the “best of both worlds.” Instead of an acoustic signal, we’re sending a laser signal into the brain, then receiving an acoustic signal back thanks to the phenomenon of the photoacoustic effect! When we’re dealing with a mixed medium like human tissue, the ultrasound signal often gets distorted by the time it’s picked up by the receiving probe. Thankfully in photoacoustics, the acoustic signal only travels one way, so compared to what happens in traditional ultrasound imaging, it goes through less distortion by the time it reaches the ultrasound probe. One example of a mixed medium is the skull, which has a very porous layer in the middle (see Figure 1).

Figure 1. Schematic of the photoacoustic imaging process. Image courtesy of Hyeonu Heo.

Over the past several years, it has been a challenge trying to get a good acoustic signal when imaging through the skull. However, as you’re probably aware, AI has lately become popular in enhancing different applications, and the biomedical field is no exception to that. This project proposes using a deep learning model to improve photoacoustic images distorted by the skull, by training it on these images and comparing them to their “ground truth” counterparts. By the time the model is finished training, it will be able to improve the quality of images that it has never seen before! The model we’re working with is called U-Net, named after its literal shape of a U. The two main parts of U-Net are the encoder on the left side and the decoder on the right side (see Video 1). The encoder takes an image, lowers its resolution, and extracts important features out of it to learn from. Later, the decoder restores the image’s resolution and is able to pinpoint down to each pixel which parts of the image are what (for example, a blood vessel or background).

So far in our research, we’ve noticed how the types of images that we feed the model are very important. For instance, if we train it on images that only consider 2D wave effects, it isn’t going to perform as well when tested on images with more realistic 3D wave effects. It is crucial that the training data for the model is as realistic as possible, before it gets deployed out into the real world to be used in clinical settings. Fortunately, our U-Net model has proven to be very robust, and the results that we’ve obtained thus far have pointed us toward ways to further improve it. The future in this field is exciting, since several categories of photoacoustic imaging tasks can benefit from deep learning enhancements, such as monitoring stroke diseases.